Advances in Neural Networks - ISNN 2006

This paper presents an effective pattern classification model by designing an artificial neural network based pattern classifiers for face recognition. First, a RGB image inputted from a frame grabber is converted into a HSV image. Then, the coarse facial region is extracted using the hue(H) and saturation(S) components except intensity(V) component which is sensitive to the environmental illumination. Next, the fine facial region extraction process is performed by matching with the edge and gray based templates. To make a light-invariant and qualified facial image, histogram equalization and intensity compensation processing using illumination plane are performed. The finally extracted and enhanced facial images are used for training the pattern classification models. The proposed hierarchical ART2 pattern classification model which has the Max-Min cluster selection strategy makes it possible to search clustered reference patterns effectively. Experimental results show that the proposed face recognition system is as good as the SVM model which is famous for face recognition field in recognition rate and even better in classification speed.

When training set is unbalanced, the conventional least square error (LSE) training strategy is less efficient to train neural network (NN) for classification because it often lead the NN to overcompensate for the dominant group. Therefore, in this paper a dynamic threshold learning algorithm (DTLA) is proposed as the substitute for the conventional LSE algorithm. This method uses multiple dynamic threshold parameters to gradually remove some training patterns that can be classified correctly by current Radial Basis Function (RBF) network out of the training set during training process, which changes the unbalanced training problem into a balanced training problem and improves the classification rate of the small group. Moreover, we use the dynamical threshold learning algorithm to classify the remote sensing images, when the unbalanced level of classes is high, a good effect is obtained.

Constructive fuzzy neural networks (i.e., CFNN) proposed in [1] cannot be used for non-numerical data. In order to use CFNN to deal with non-numerical complicated data, rough set theory is adopted to improve the CFNN in this paper. First of all, we use rough set theory to extract core set of non-numerical attributes and decrease number of dimension of samples by reducing redundancy. Secondly, we can pre-classify the samples according to non-numerical attributes. Thirdly, we use CFNN to classify the samples according to numerical attributes. The proposed method not only increases classification accuracy but also speeds up classification process. Finally, the classification of wireless communication signals is given as an example to illustrate the validation of the proposed method in this paper.

This paper is focused on cellular phone embedded eye location system. The proposed eye detection system is based on a hierarchy cascade FloatBoost classifier combined with an MLP neural net post classifier. The system firstly locates the face and eye candidates’ areas in the whole image by a hierarchical FloatBoost classifier. Then geometrical and relative position information of eye-pair and the face are extracted. These features are input to a MLP neural net post classier to arrive at an eye/non-eye decision. Experimental results show that our cellular phone embedded eye detection system can accurately locate double eyes with less computational and memory cost. It runs at 400ms per image of size 256×256 pixels with high detection rates on a SANYO cellular phone with ARM926EJ-S processor that lacks floating-point hardware.

In this paper, we discuss human iris recognition, which is based on iris localization, feature extraction, and classification. The features for iris recognition are extracted from the segmented iris pattern using two-dimensional (2-D) wavelet transform based on Haar wavelet. We present an efficient initialization method of the weight vectors and a new method to determine the winner in LVQ neural network. The proposed methods have more accuracy than the conventional techniques.

In this paper, a novel iris recognition method is proposed based on a state-of-the-art classification technique called minimax probability machine (MPM). Engaging the binary MPM technique, this work develops a multi-class MPM classification for reliable iris recognition with high accuracy. The experiments on iris database demonstrate that compared to the existent methods, the MPM-based iris recognition algorithm obtains better classification performance. It can significantly improve the recognition accuracy and has a competitive and promising performance.

In this paper, we present an efficient algorithm of extracting the multiple facial features such as eyes, nose, and mouth. The face candidates are first obtained based on skin-color filtering in

YC

b

C

r

color domain and skin-temperature values and then the elliptic measures are applied to extract a true face candidate and its boundary. A Sobel edge mask is performed and consequently horizontal projection operation is applied to locate the eyes referring to the maximum horizontal projection value in

Y

component. Once two eyes are located, the distance that crosses the center of eyes and extends to the face boundary,

D

1

is determined. A heteroassociative memory neural network model is utilized to find the facial features. An input neuron vector

X

accepts

D

1

and the output neurons vector

Y

maps it to the facial features such as eyes, nose and mouth.

PCA and NMF subspace approaches have become the most representative methods in face recognition, which act in the similar way as a neural network auto-associative memory. By integrating with

LDA

subspace, in this paper, two subspace associative memories,

PCA

LDA

and

NMF

LDA

, are proposed, and how they recognize the partially damaged faces is presented. The theoretical expressions are plotted, and the comparative experiments are completed for the UMIST face database. It shows that

NMF

LDA

subspace associative memory outperform

PCA

LDA

subspace method significantly in recognizing partially damaged faces.

In this paper, we try to develop an analytical framework for classifying human basic emotions. We try to find out what are the major components of each facial expression, what are the patterns that distinguish them from one another. We applied widely used pattern recognition technique-principle component analysis to characterize the feature point displacements of each basic human facial expression for each individual in the existing database. For faces not existent in the database, so called “novel face” in our experiment, we will first find the face in the database that has most likely neutral face to this individual, and base on an assumption that are widely accepted in cognitive science, we will classify this novel face to the category where the most similar one belongs, and classifying his/her facial expression using the so called “expression model” of the most similar individual. This kind of approach has never be exploited before, then we will examine its robustness in our experiment.

We present a fully automatic real time system for face detection and basic facial expression recognition from video and images. The system automatically detects frontal faces in the video stream or images and classifies each of them into 7 expressions. Each video frame is first scanned in real time to detect upright-frontal faces. The faces found are scaled into image patches of equal size and sent downstream for further processing. Gabor energy filters are applied at the scaled image patches followed by a recognition engine. Best results are obtained by selecting a subset of Gabor features using AdaBoost and then training Support Vector Machines on the outputs of the features selected by AdaBoost.

This paper describes a facial expression recognition system based upon Active Appearance Model (AAM), which has been typically used for the face recognition task. Given that AAM has been also used in tracking the moving object, we thought it could be effective in recognizing the facial expressions of humans. Our results show that the performance of the facial expression recognition using AAM is reliably high when it combined with an enhanced Fisher classification model.

In recent years, facial expression recognition has become an active research area that finds potential applications in the fields such as images processing and pattern recognition, and it plays a very important role in the applications of human-computer interfaces and human emotion analysis. This paper proposes an algorithm called BoostingTree, which is based on the conventional Adaboost and uses tree-structure to convert seven facial expressions to six binary problems, and also presents a novel method to compute projection matrix based on Principal Component Analysis (PCA). In this novel method, a block-merger combination is designed to solve the “data disaster” problem due to the combination of eigenvectors. In the experiment, we construct the weak classifiers set based on this novel method. The weak classifiers selected from the above set by Adaboost are combined into strong classifier to be as node classifier of one level of the tree structure. N-level tree structure built by BoostingTree can effectively solve multiclass problem such as facial expression recognition

Kernel discriminant analysis (KDA) and the kernel principal component analysis (KPCA), which are the extension of the linear discriminant analysis (LDA) and the principal component analysis (PCA), respectively, from linear domain to nonlinear domain via the kernel trick, are two very popular nonlinear feature extraction methods. In this paper, we present a new feature extraction algorithm by combing KDA and KPCA, and then apply it to the face recognition task. The experimental results on Yale face dataset show that the proposed method can significantly improve the performance both KDA and KPCA.

A novel chaotic neural network K-set has been constructed based in research on biological olfactory systems. This non-convergent neural network simulates the capacities of biological brains for signal processing in pattern recognition. Its accuracy and efficiency are demonstrated in this report on an application to human face recognition, with comparisons of performance with conventional pattern recognition algorithms.

The human ability to recognize objects has not so far been matched by intelligent machines. This is more evident when it comes to recognizing faces, where a quick human “glance” is sufficient to recognize a “familiar” face. Face recognition has recently attracted more research aimed at developing reliable recognition by machines. Current face recognition methods rely on detecting certain features within a face and using these features for face recognition. This paper introduces a novel approach to face recognition by simulating our ability to recognize “familiar” faces after a quick “glance” using pattern averaging and neural networks. A real-life application will be presented throughout recognizing the faces of 30 persons. Time costs and the neural network parameters will be described, in addition to future work aimed at further improving the developed system.

Recently semi-supervised learning has attracted a lot of attention. Different from traditional supervised learning, semi-supervised learning makes use of both labeled and unlabeled data. In face recognition, collecting labeled examples costs human effort, while vast amounts of unlabeled data are often readily available and offer some additional information. In this paper, based on Support Vector Machine (SVM), we introduce a novel semi-supervised learning method for face recognition. The basic idea of the method is that, if two data points are close to each other, they tend to share the same label. Therefore, it is reasonable to search a projection with maximal margin and locality preserving property. We compare our method to standard SVM and transductive SVM. Experimental results show efficiency and effectiveness of our method.

This paper proposes a method for face recognition by integrating non-negative matrix factorization with sparseness constraints (NMFs) and radial basis function (RBF) classifier. NMFs can represent a facial image based on either local or holistic features by constraining the sparseness of the basis images. The comparative experiments are carried out between NMFs with low or high sparseness and principle component analysis (PCA) for recognizing faces with or without occlusions. The simulation results show that RBF classifier outperforms

k

–nearest neighbor linear classifier significantly in recognizing faces with occlusions, and the holistic representations are generally less sensitive to occlusions or noise than parts-based representations.

In this paper, we propose a face detection method by combining classifiers. We apply two classifiers using features extracted from complementary feature subspaces learned by principal component analysis (PCA). The two classifiers employ the same classification model named a polynomial neural network (PNN). The outputs of the two classifiers are fused to make the final decision. The effectiveness of the proposed method has been demonstrated in experimentals.

A face detection method based on Kernel Independent Component Analysis and Boosting Chain Algorithm was proposed. Moreover a linear optimization scheme was proposed to address the problems of redundancy in boosting learning and threshold adjusting in cascade coupling. Experiments were done to compare the performance of boosting chain with that of Adaboost and Floatboost and the results show the effectiveness of this new method.

One of the main challenges faced by the current face recognition techniques lies in the difficulties of collecting samples, and many existing face recognition techniques rely heavily on the size and representative of training set. Those algorithms may suffer serious performance drop or even fail to work if

only one training sample per person

is available to the systems. In this paper, we present a multiple-SOMs-based fusion method to address this problem. Based on the localization of the face, multiple Self-Organizing Maps are constructed in different manners, and then fused to obtain a more compact and robust representation of the face, through which the discrimination and class-specific information can be easily explored from the single training image among a large number of classes. Experiments on the FERET face database show that the proposed fusion method can significantly improve the performance of the recognition system, achieving a top 1 matching rate of 90.0%.

This paper introduces a new concept of LLE eigenface modelled by local linear embedding (LLE), and compares it with the traditional PCA eigenface from principle component analysis (PCA) on pose identity and face identity recognition through face classification. LLE eigenface is found outperforming PCA eigenface on the discrimination/recogntion of both face identity and pose identity. The superiority on face identity recognition is own to a more balanced energy distribution on LLE eigenfaces, while the superiority on pose identity recognition is due to the fact that LLE preserves a better local neighborhood of face images.

In this paper, we propose a priority verification method for multimodal biometric features by using a momentum back-propagation artificial neural network (MBP-ANN). We also propose a personal verification method using both face and speech to improve the rate of single biometric verification. False acceptance rate (FAR) and false rejection rate (FRR) have been a fundamental bottleneck of real-time personal verification. The proposed multimodal biometric method is to improve both verification rate and reliability in real-time by overcoming technical limitations of single biometric verification methods. The proposed method uses principal component analysis (PCA) for face recognition and hidden markov model (HMM) for speech recognition. It also uses MBP-ANN for the final decision of personal verification. Based on experimental results, the proposed system can reduce FAR down to 0.0001%, which proves that the proposed method overcomes the limitation of single biometric system and proves stable personal verification in real-time.

This paper first introduces the system requirement and the system flow of the auto-plotting system. As the data points needed by the auto-plotting system coming from the remote speech signals, to reach high recognition accuracy, the Hidden Markov Model (HMM) approach was chosen as the speech recognition approach. Then the paper is detailed on the speaker dependent (SD), speaker independent (SI) and speaker adaptive (SA) speech recognition methods. We proposed the n-speech models SD system as the recognition system to gain the highest recognition performance in varying speech environments. However the system required that searching for the optimal model from the database should finish in 5 minutes, so the paper finally describes how the Self-Organizing Map (SOM) was used to pre clustering to the n-speech models, to decrease the time for speech recognition and results evaluation and decrease matching time, Experiments show the n-speech models SD system can select the best-matching model in the limited time and improve the average speech recognition accuracy to 97.2. It ideally suits the system requirements.

We studied the application of Biomimetic Pattern Recognition to speaker recognition. A speaker recognition neural network using

network matching degree

as criterion is proposed. It has been used in the system of text-dependent speaker recognition. Experimental results show that good effect could be obtained even with lesser samples. Furthermore, the misrecognition caused by untrained speakers occurring in testing could be controlled effectively. In addition, the basic idea “cognition” of Biomimetic Pattern Recognition results in no requirement of retraining the old system for enrolling new speakers.

In this paper, we propose a new text-independent speaker identification method using VQ and MLP. It consists of three parts: a new spectral peak analysis based feature extraction, speaker clustering and model selection using VQ, and MLP based speaker identification. The feature vector reflects the speaker specific characteristics and has a long-term feature for which makes it text-independent. The proposed method has a computational efficient for feature extraction and identification. To evaluate the proposed method, we calculated the correct identification ratio (CIR), the average CIR of the proposed and GMM method was 92.27% and 85.78% for 5 seconds segments in 15-speaker identification. Experimental results, we have achieved a performance comparable to GMM-method.

In this paper, the neural net pattern recognition equations were attempted to apply to speech recognition. The proposed method features a dynamic process of self-organization that has been proved to be successful in recognizing a depth perception in stereoscopic vision. This study showed that the dynamic process was also useful in recognizing human speech. In the processing, input vocal signals are first compared with standard models to measure similarities that are then given to the dynamic process of self-organization. The competitive and cooperative processes are conducted among neighboring input similarities, so that only one winner neuron is finally detected. In a comparative study, it showed that the proposed method outperformed the conventional Hidden Markov Models(HMM) speech recognizer under the same conditions.

A method is proposed for classifying music genre for audio retrieval systems using time-delay neural networks. The proposed classification method considers eight types of music genre: Blues, Country, Hard Core, Hard Rock, Jazz, R&B(Soul), Techno, and Trash Metal. The melody between bars in the music is used to distinguish the different genres. The melody pattern is extracted based on the sound of a snare drum, which is used to effectively represent the rhythm periodicity. Classification is based on a time-delay neural network that uses a Fourier transformed vector of the melody as an input pattern. This classification method was used to analyze 80 training data from ten different musical pieces for each genre and a further 40 test data from five additional musical pieces for each genre. The accuracy of the genre classifications that were obtained for the two sets of data was 92.5% and 60%, respectively.

As the internet community grows larger, digital music distribution becomes widely available and is made easier than ever. Artists from all over the world can make their songs available by a single click. Websites, containing varieties of music style for download, charge only a fraction of the cost of a CD for the service. With the incredible amount of music pieces available, it is impossible to classify each piece by its style manually. A procedure is proposed using the support vector statistical learning algorithm to achieve the task autonomously. Digital music files are converted, partitioned and processed to obtain the desirable input vectors for the algorithm. As the machine learns the features of each music genre, it is capable of classifying input vectors from unknown pieces. A simulation was carried out to evaluate the efficiency of the algorithm. Results from the simulation are presented and discussed in this paper. Conclusions are drawn by comparing other algorithms against the proposed method.

This paper presents a novel approach for gender classification by boosting local binary pattern-based classifiers. The face area is scanned with scalable small windows from which Local Binary Pattern (LBP) histograms are obtained to effectively express the local feature of a face image. The Chi square distance between corresponding Local Binary Pattern histograms of sample image and template is used to construct weak classifiers pool. Adaboost algorithm is applied to build the final strong classifiers by selecting and combining the most useful weak classifiers. In addition, two experiments are made for classifying gender based on local binary pattern. The male and female images set are collected from FERET databases. In the first experiment, the features are extracted by LBP histograms from fixed sub windows. The second experiment is tested on our boosting LBP based method. Finally, the results of two experiments show that

the features extracted by LBP operator are discriminative for gender classification

and our proposed approach achieves better performance of classification than several others methods.

In this paper, we present a novel approach to multi-view gender classification considering both shape and texture information to represent facial image. The face area is divided into small regions, from which local binary pattern(LBP) histograms are extracted and concatenated into a single vector efficiently representing the facial image. The classification is performed by using support vector machines(SVMs), which had been shown to be superior to traditional pattern classifiers in gender classification problem. The experiments clearly show the superiority of the proposed method over support gray faces on the CAS-PEAL face database and a highest correct classification rate of 96.75% is obtained. In addition, the simplicity of the proposed method leads to very fast feature extraction, and the regional histograms and global description of the face allow for multi-view gender classification.

Through task decomposition and module combination, min-max modular support vector machines (M

3

-SVMs) can be successfully used for different pattern classification tasks. Based on an equal clustering algorithm, M

3

-SVMs can divide the training data set of the original problem into several subsets with nearly equal number of samples, and combine them to a series of balanced subproblems which can be trained more efficiently and effectively. In this paper, we explore the use of M

3

-SVMs with equal clustering method in gender recognition. The experimental results show that M

3

-SVMs with equal clustering method can be successfully used for gender recognition and make the classification more efficient and accurate.

This paper proposes a novel method for recognizing palmprint using the winner-take-all (WTA) network based independent component analysis (ICA) algorithm and the radial basis probabilistic neural network (RBPNN) proposed by us. The WTA-ICA algorithm exploits the maximization of the sparse measure criterion as the cost function, and it extracts successfully palmprint features. The classification performance is implemented by the RBPNN. The RBPNN is trained by the orthogonal least square (OLS) algorithm and its structure is optimized by the recursive OLS (ROLS) algorithm. Experimental results show that the RBPNN achieves higher recognition rate and better classification efficiency with other usual classifiers.

A traditional studies about recognition and representation technology of sign language have several restrictions such as conditionality in space and limitation of motion according to the technology of wire communication, problem of image capture system or video processing system for an acquisition of sign language signals, and the sign language recognition system based on word and morpheme. In order to overcome these restrictions and problems, in this paper, we implement the Korean sign language recognizer in the shape of sentence using neural network based on the Post wearable PC platform. The advantages of our approach are as follows: 1) it improves efficiency of the sign language input module according to the technology of wireless communication, 2) it recognizes and represents continuous sign language of users with flexibility in real time, and 3) it is possible more effective and free interchange of ideas and information between deaf person and hearing person (the public). Experimental result shows the average recognition rate of 92.8% about significant, dynamic and continuous the Korean sign language.

This paper proposes an approach to automatic gait recognition based on wavelet descriptors and independent component analysis (ICA) for the purpose of human identification at a distance. Firstly, the background extraction method is applied to subtract the moving human figures accurately and to obtain binary silhouettes. Secondly, these silhouettes are described with wavelet descriptors and converted into one-dimensional signals to get the independent components (ICs) of these feature signals through ICA. Then, a fast and robust fixed-point algorithm for calculating the ICs is adopted and a selection criterion how to choose ICs is given. Lastly, the nearest neighbor and support vector machine classifiers are chosen for recognition and the method is tested on the XAUT and NLPR gait database. Experimental results show that our method has encouraging recognition accuracy with comparatively low computational cost.

This paper presents a new method for human model-free gait recognition using principal curves analysis and neural networks. Principal curves are non-parametric, nonlinear generalizations of principal component analysis, and give a breakthrough to nonlinear principal component analysis. Different from the traditional statistical analysis methods, principal curve analysis seeks lower-dimensional manifolds for every class respectively, and forms the nonlinear summarization of the sample features and directions for each class. Neural network with the virtue of its universal approximation property is an outstanding method to model the nonlinear function of principal curve. Firstly, a background subtraction is used to separate objects from background. Secondly, we extract the contour of silhouettes and represent the spatio-temporal features. Finally, we use principal curves and neural networks to analyze the features to train and test gait sequences. Recognition results demonstrate that our method has encouraging recognition performance.

This paper presents a method to detect adjacent multiple pedestrians using the ART2 neural network from a moving camera image. A BMA(Block Matching Algorithm) is used to obtain a motion vector from two consecutive input frames. And a frame difference image is generated by the motion compensation with the motion vector. This image is transformed into binary image by the adapted threshold and a noise is also removed. To detect multiple pedestrians, a projection histogram is processed by the shape information of human being. However, in case that pedestrians exist adjacently each other, it is very different to separate them. So, we detect adjacent multiple pedestrians using the ART2 neural network. The experimental results on our test sequences will show the high efficiency of our method.

A novel recognition method of throwing force of athlete combined with wavelet and multi-class support vector machine is introduced in the paper, which is based on the analysis of motion characters of gliding shot put. Utilizing the digital shot based on a three dimensional accelerometer, we get the three dimensional throwing forces in real time. Through wavelet transform, the general characteristics of force information are picked up. Then the general characteristics are input into the classifier for recognition of throwing force curves. The analysis provides the scientific basis for the motion training and instruction of shot put. The experiment shows that the method not only has high anti-noise ability and improves the recognition efficiency, but also decreases the burden of system and improves the recognition speed.

The paper presents a new constructive learning algorithm CWSN (Covering With Sphere Neighborhoods) for three-layer neural networks, and uses it to solve the text categorization (TC) problem. The algorithm is based on a geometrical representation of M-P neuron, i.e., for each category, CWSN tries to find a set of sphere neighborhoods which cover as many positive documents as possible, and don’t cover any negative documents. Each sphere neighborhood represents a covering area in the vector space and it also corresponds to a hidden neuron in the network. The experimental results show that CWSN demonstrates promising performance compared to other commonly used TC classifiers.

Many applications, such as word-sense disambiguation and information retrieval, can benefit from text classification. Text classifiers based on Independent Component Analysis (ICA) try to make the most of the independent components of text documents and give in many cases good classification effects. Short-text documents, however, usually have little overlap in their feature terms and, in this case, ICA can not work well. Our aim is to solve the short-text problem in text classification by using Latent Semantic Analysis (LSA) as a data preprocessing method, then employing ICA for the preprocessed data. The experiment shows that using ICA and LSA together rather than only using ICA in Chinese short-text classification can provide better classification effects.

This paper describes the design and implementation of a system that identify the writer using off-line Arabic handwriting. Our approach is based on the combination of global and structural features. We used genetic algorithm for feature subset selection in order to eliminate the redundant and irrelevant ones. A modular Multilayer Perceptron (MLP) classifier was used. Experiments have shown writer identification accuracies reach acceptable performance levels with an average rate of 94.73% using optimal feature subset. Experiments are carried on a database of 180 text samples, whose text was made to ensure the involvement of the various internal shapes and letters locations within a word.

In this paper, we present a novel self-generation, supervised character recognition algorithm based on adaptive resonance theory (ART) artificial neural network (ANN) and delta-bar-delta method. By combining two methods, the proposed algorithm can reduce noise problem in the ART ANN and the local minima problem in the delta-bar-delta method. The proposed method can extend itself based on new information contained in input patterns that require nodes of hidden layers in neural networks and effectively find characters. We experiment with various real-world documents such as a student ID and an identifier on a container. The experimental results show that the proposed self-generation. ART algorithm reduces the possibility of local minima and accelerates learning speed compared with existing.

A novel approach for handwritten digit recognition is proposed in this paper, which combines the low rank approximation and the competitive neural network together. The images in each class are clustered into several subclasses by the competitive neural network, which is helpful for feature extraction. The low rank approximation is used for image feature extraction. Finally, the k-nearest neighbor classifier is applied to the classification. Experiment results on USPS dataset show the effectiveness of the proposed approach.

Pattern recognition is a well-established field of study and Optical Character Recognition (OCR) has long been seen as one of its important contributions. However, Arabic has been one of the last major languages to receive attention. This paper describes the performance of an approach combining Hough transform in features extraction and Neural Networks in classification. Experimental tests have been carried out on a set of 85.000 samples of characters corresponding to5 different fonts. Some promising experimental results are reported.

A novel hierarchical algorithm is proposed to recognize English calling cards. The algorithm processes multiresolution images of calling cards hierarchically to firstly extract individual characters and then to recognize the characters by using an enhanced neural network method. The horizontal smearing is applied to a 1/3 resolution image in order to extract the areas. The second vertical smearing and contour tracking masking is applied to a 1/2 resolution image to extract individual characters. And lastly, the original image is used in the recognition step because the image accurately includes the morphological information of the characters precisely. The enhanced RBF network is also proposed to recognize characters with diverse font types and sizes, by using the enhanced ART1 network adjusting the vigilance parameter dynamically according to the similarity between patterns. The results of experiments show that the proposed algorithm greatly improves the character extraction and recognition compared with traditional recognition algorithms.

A professional Chinese fax information processing system is designed which has functions to automate incoming fax distribution in a company or institution, read an incoming fax cover sheet and route the fax to the receiver’s email box. This paper reports our research as part of an effort to realize such a system and focuses on recognition of the handwritten recipient’s on fax cover pages. We propose hybrid neural networks for large scale Chinese handwritten character recognition. The network is composed of the self-organizing competitive fuzzy layer and the multi-layer neural network using BP method, connected in cascade. The characteristic features of this network structure for Chinese handwritten character recognition are discussed and performances are evaluated on 8208 real world faxes which are taken from one company in 2004, the results of experiments compared to standard neural solutions based on MLP show that the whole system is of reasonable structure and satisfactory performance.

Photo time-stamp is a valuable information source for content-based retrieval of scanned photo databases. A fast photo-stamp recognizing approach based on Self-Generating Neural Networks (SGNN) is proposed in this paper. Network structures and parameters of SGNN needn’t to be set by users, and their learning process needn’t iteration, so SGNN can be trained on-line. Proposed method consists of three steps: A photo is roughly segmented to determine which corner of the photo contains time-stamp; The area which contains time-stamp of the photo is finely segmented, in order to locate each character in the time-stamp, projection technology is used to locate edges of these characters; The time-stamp is recognized based on SGNN. Experimental results show that proposed approach can achieve higher recognition accuracy and computing efficiency.

We propose a hierarchical classifier of object images using neural networks for content-based image classification. The images for classification are object images that can be divided into foreground and background areas. In the preprocessing step, we extract the object region and shape-based texture features extracted from wavelet-transformed images. We group the image classes into clusters that have similar texture features using Principal Component Analysis (PCA) and K-means. The hierarchical classifier has five layers that combine the clusters. The hierarchical classifier consists of 59 neural network classifiers that were learned using the back-propagation algorithm. Of the various texture features, the diagonal moment was the most effective. A test showed classification rates of 81.5% correct with 1000 training images and of 75.1% correct with 1000 test images. The training and test sets each contained 10 images from each of 100 classes.

Image classification is a challenging problem in organizing a large image database. However, an effective method for such an objective is still under investigation. This paper presents a method based on wavelet for image classification with adaptive processing of data structures. After decomposed by wavelet, the features of an image can be reflected by the wavelet coefficients. Therefore, the nodes of tree representation of images are represented by distribution of histograms of wavelet coefficient projections. 2940 images derived from seven original categories are used in experiments. Half of the images are used for training neural network and the other images used for testing. The classification rate of training set is 90%, and the classification rate of testing set is 87%. The promising results prove the proposed method is very effective and reliable for image classification.

Considering the feature of remote sensing images, we put forward a remote sensing image classification algorithm based on Hopfield neural network. First, the function and principle of Hopfield neural network is described in this paper. Then based on the common model of Hopfield neural network, the image classification algorithm using Hopfield neural network is realized and experimental results show that its precision is superior to that of the conventional maximum likelihood classification algorithm.

Based on the research on mechanism of biological olfactory system, we constructed a K-set, which is a novel bionic neural network. Founded on the groundwork of K0, KI and KII sets, the KIII set in the K-set hierarchy simulates the whole olfactory neural system. In contrast to the conventional artificial neural networks, the KIII set operates in nonconvergent ‘chaotic’ dynamical modes similar to the biological olfactory system. In this paper, an application of electronic nose-brain for tea classification using the KIII set is presented and its performance is evaluated in comparison with other methods.

A new algorithm to distinguish onion leaves from weed leaves in images is suggested. This algorithm is based on segmentation of color images and on BP neural network. It includes: discarding soil for conserving only plants in the image, color image segmentation, merging small regions by analyzing the frontier rates and the averages of color indices of the regions, at last a BP neural network is used to determine if the small regions belongs to onion leaf or not. The algorithm has been applied to many images and the correct identifiable percents for onion leaves are between 80%~ 90%.

In this paper, a new method for bark classification based on textural and fractal dimension features using Artificial Neural Networks is presented. The approach involving the grey level co-occurrence matrices and fractal dimension is used for bark image analysis, which improves the accuracy of bark image classification by combining fractal dimension feature and structural texture features on bark image. Furthermore, we have investigated the relation between Artificial Neural Network (ANN) topologies and bark classification accuracy. Furthermore, the experimental results show the facts that this new approach can automaticly identify the Tplants categories and the classification accuracy of the new method is better than that of the method using the nearest neighbor classifier.

This paper presents a fast neural network method of radial basis function with dynamic decay adjustment (RBFN-DDA) to classify Quasi-Stellar Objects (QSOs) and galaxies automatically. The classification process is mainly comprised of three parts: (1) the dimensions of the normalized input spectra is reduced by the Principal Component Analysis (PCA); (2) the network is built from scratch: the number of required hidden units is determined during training and the individual radii of the Gaussians are adjusted dynamically until corresponding criterions are satisfied; (3) The trained network is used for the classification of the real spectra of QSOs and galaxies. The method of RBFN-DDA having constructive and fast training process solves the difficulty of selecting appropriate number of neurons before training in many methods of neural networks and achieves lower error rates of spectral classification. Besides, due to its efficiency, the proposed method would be particularly useful for the fast and automatic processing of voluminous spectra to be produced from the large-scale sky survey project.

The feed-forward neural network using simulated annealing resilient propagation (SARPROP) algorithm was applied to the research community of radar target recognition in this paper. The high resolution radar range profiles were selected as the feature vectors for data representation, and the product spectrum based features were introduced to improve classification performance. Simulations are presented to identify the four different aircrafts. The results show that the SARPROP algorithm combined with product spectrum based features is effective and robust for the application of radar target recognition.

In this paper, RBF network (RBFN) is used to provide effective methodologies for solving difficult computational problems in camera calibration and 3D reconstruction process. RBFN works in three aspects: Firstly, a RBFN is adopted to learn and memorize the nonlinear relationship in stereovision system. Secondly, another RBFN is trained to search the correspondent lines in two images such that stereo matching is performed in one dimension. Finally, the trained network in the first stage is used to reconstruct the object’s 3D figuration and surface. The technique avoids the complicated and large calculation in conventional methods. Experiments have been performed on common stereo pairs and the results are accurate and convincing.

This study describes a full model of calibrating an omnidirectional stereo vision system, which includes the rotation and translation between the camera and mirrors, and an algorithm implemented with a backpropagation technique of the neural network to determine this relative position from observations of known points in a single image. The system is composed of a perspective camera and two hyperbolic mirrors, which are configured to be separate and coaxial besides sharing one focus that coincides with the camera center for providing a single projection point. We divide the calibration into two steps. The first step we calibrate the camera’s intrinsics without the mirrors in order to reduce computational complexity and in the second step we estimate the pose parameters of the CCD camera with respect to the mirrors based on a Levenberg-Marquart Backpropagation (LMBP) algorithm. The proposed tech- nique can be easily applied to all kinds of catadioptric sensors and various amounts of misalignment between the mirrors and cameras.

This paper presents an evolutionary cellular automata based neural systems (Evolutionary CANS) for visual servoing of RV-M2 robot manipulator. The architecture of CANS consist of a two-dimensional (2-D) array of basic neurons. Each neuron of CANS has local connections only with contiguous neuron and acts as a form of pulse according to the dynamics of the chaotic neuron model. CANS are generated from initial cells according to the cellular automata (CA) rule. Therefore neural architecture is determined by both initial pattern of cells and production rule of CA. Production rules of CA are evolved based on a DNA coding. DNA coding has the redundancy and overlapping of gene and is apt for representation of the rule. In this paper we show the general expression of CA rule and propose translating method from DNA code to CA rule. In addition, we present visual servoing application using evolutionary CANS.

Robust visual object tracking is one of the key problems in computer vision. Subspace based tracking method is a promising approach in handling appearance variability. Linear Discriminant Analysis(LDA) has been applied to this problem, but LDA is not a stable algorithm especially for visual tracking. Maximum Margin Criterion(MMC) is a recently proposed discriminant criterion. Its promising specialities make it a better choice for the tracking problem. In this paper, we present a novel subspace tracking algorithm based on MMC. We also proposed an incremental version of the corresponding algorithm so that the tracker can update in realtime. Experiments show our tracking algorithm is able to track objects well under large lighting, pose and expression variation.

In this paper an attention selection system based on neural network is proposed, which combines supervised and unsupervised learning reasonably. A value system and memory tree with update ability are regarded as teachers to adjust the weights of neural network. Both bottom-up and top-down part are to simulate two-stage hypothesis of attention selection in biological vision. The system is able to track objects that it is interested in. Whenever it lost focus on tracked object, it can find the object again in a short time.

This paper presents a method of human motion tracking based on Markov random field and Hopfield neural networks. The model of rigid body motion is first introduced in the MRF-based motion segmentation. The potential function in MRF is defined according to this motion model. The Hopfield neural network is first used in the implementation of MRF to take advantage of some mature Neural Network technique. After the introduction of the model of rigid body motion the joint angles of human body can be estimated .It is also helpful to the estimation of the proportions of human body, which is significant to the accurate estimation of human motion. Finally the experimental results are given and the existed problems in this method are pointed out.

We develop a simple and fast human tracking system based on skin-color using Kalman filter for humanoid robots. For our human tracking system we propose a fuzzy and probabilistic model of observation noise, which is important in Kalman filter implementation. The uncertainty of the observed candidate region is estimated by neural network. Neural network is also used for the verification of face-like regions obtained from skin-color information. Then the probability of observation noise is controlled based on the uncertainty value of the observation. Through the real-human tracking experiments we compare the performance of the proposed model with the conventional Gaussian noise model. The experimental results show that the proposed model enhances the tracking performance and also can compensate the biased estimations of the baseline system.

We present a framework for object detection via fusion of global classifier and part-based classifier in this paper. The global classifier is built using a boosting cascade to eliminate most non-objects in the image and give a probabilistic confidence for the final fusion. In constructing the part-based classifier, we boost several neural networks to select the most effective object parts and combine the weak classifiers effectively. The fusion of these two classifiers generates a more powerful detector either on efficiency or accuracy. Our approach is evaluated on a database of real-world images containing rear-view cars. The fused classifier gives distinctively superior performance than traditional cascade classifiers.

By analyzing the particular features of visual content for cartoon videos, 8 typical features of MPEG-7 descriptors are extracted to distinguish the cartoons from other videos. Then, a content-based video classifier is developed by combining the active relevance feedback technique and SVM for detecting the cartoon videos. The experimental results on the vast real video clips illustrate that compared with the classifier based on SVM and that based on traditional relevance feedback technique and SVM, the proposed classifier has a higher advantage of cartoon video detection.

An effective morphological neural network of background clutter prediction for detecting small targets in image data is proposed in this paper. The target of interest is assumed to have a very small spatial spread, and is obscured by heavy background clutter. The clutter is predicted exactly by morphological neural networks and subtracted from the input signal, leaving components of the target signal in the residual noise. Computer simulations of real infrared data show better performance compared with other traditional methods.

In this paper, an effective classification approach for action scenes is proposed, which exploits the film grammar used by filmmakers as guideline to extract features, detect and classify action scenes. First, action scenes are detected by analyzing film rhythm of video sequence. Then four important features are extracted to characterize chase and fight scenes. After then the Probability Neural Networks is employed to classify the detected action scenes into fight, chase and uncertain scenes. Experimental results show that the proposed method works well over the real movie videos.

In this paper, we study sparse representation of large-size natural scenes via local spatial dependency decomposition. We propose a local independent factorization model of natural scenes and develop a learning algorithm for adaptation of the synaptic weights. We investigate the dependency of neighboring location of the natural scene patches and derive learning algorithm to train the visual neural network. Some numerical experiments on natural scenes are performed to show the sparse representation of the visual sensory information.

We present a new search region prediction method using frequency sensitive competitive learning vector quantization for motion estimation of image sequences. The proposed method can decrease the computation time because of the smaller number of search points compared to other methods, and also reduces the bits required to represent motion vectors. The results of experiments show that the proposed method provides competitive PSNR values compared to other block matching algorithms while reducing the number of search points and minimizing the complexity of the search region prediction process.

A new extraction method for remote sensing data is proposed by using both a support vector machine (SVM) and knowledge reasoning technique. The new method fulfils intelligent extraction of water, road and other plane-like objects from remote sensing images in a hierarchical manner. It firstly extracts water and road information by a SVM and pixel-based knowledge post-processing method, then removes them from original image, and then segments other plane-like objects using the SVM model and computes their features such as texture, elevation, slope, shape etc., finally extracts them by the polygon-based uncertain reasoning method. Experimental results indicate that the new method outperforms the single SVM and moreover avoids the complexity of single knowledge reasoning technique.

This paper proposed a neural network based method for eyes location. In our work, face area is first located initially using an illumination invariant face skin model; Then, it is segmented by the combination of image transformation and a competitive Hopfield neural network (CHNN) and facial feature candidates such as eyes, eyebrows and mouth are obtained; Finally, eyes are located by facial features evaluation and validation, which is based on face’s geometrical structures. Experimental results show that our system performs well under not good lighting conditions.

In this paper we present a Gabor Wavelet Network, a wavelet neural network based on Gabor functions, applied to image registration. Although wavelet network is time consuming technique, we decrease computational costs by incorporating three techniques: gradient-based feature selection, Gabor filtering, and wavelet neural network. Similarity criterion is built upon analyzing intensity function with Gabor Wavelet Network, which carries out the image registration by both gradient-based and texture features.

A novel image registration scheme is proposed. In the proposed scheme, the complete isometric mapping (Isomap) is used to extract features from the image sets, and these features are input vectors of feedforward neural networks. Neural network outputs are those translation, rotation and scaling parameters with respect to reference and observed image sets. Comparative experiments for Isomap based method, the discrete cosine transform (DCT) and Zernike moment are performed. The results show that the proposed scheme is not only accurate but also remarkably robust to noise.

As for unsupervised image segmentation, one important application is content based image retrieval. In this context, the key problem is to automatically determine the number of regions(i.e., clusters) for each image so that we can then perform a query on the region of interest. This paper presents an iterative entropy regularized likelihood (ERL) learning algorithm for cluster analysis based on a mixture model to solve this problem. Several experiments have demonstrated that the iterative ERL learning algorithm can automatically detect the number of regions in a image and outperforms the generalized competitive clustering.

Image segmentation is a long existing problem and still regarded as unsolved to a large extent in computer vision. This letter describes the modeling method of competitive neural networks and elucidates its connection with the Hopfield type optimization network. A new algorithm to map the image segmentation problem onto competitive networks is proposed and its convergence is shown by the stability analysis. Finally, the improvement on the competitive neural networks based method is validated by the simulation results.

In this paper, we present an adaptive spatially-constrained deterministic annealing (ASDA) algorithm, which takes into account the spatial continuity constraints by using a dissimilarity index that allows spatial interactions between image pixels, for image segmentation. The local spatial continuity constraint reduces the noise effect and the classification ambiguity. More importantly, the strength of spatial constraint for each given image pixel is auto-selected by the scaled variance of its neighbor pixels, which results in the adaptiveness of the presented algorithm. The effectiveness and efficiency of the presented method for image segmentation are supported by experimental results on synthetic and MR images.

In this paper, we present a new scheme to segment a given image. This scheme utilizes neuro-fuzzy system to derive a proper set of contour pixels based on multi-scale images. We use these fuzzy derivatives to develop a new curve evolution model. The model automatically detect smooth boundaries, scaling the energy term, and change of topology according to the extracted contour pixels set. We present the numerical implementation and the experimental results based on the semi-implicit method. Experimental results show that one can obtains a high quality edge contour.

Automated segmentation of MR images is a difficult problem due to the complexity of the images. Up to now, several approaches have been proposed based on spectral characteristics of MR images, but they are sensitive to the noise contained in the MR images. In this paper, we propose a robust method for noisy MR image segmentation. We use region-based features for a robust segmentation and use principle component analysis (PCA) to reduce the large dimensionality of feature space. Experimental results show that the proposed method is very tolerant to the noise and the segmentation performance is significantly improved.

The self-localization problem is very important when the mobile robot has to move in autonomous way. Among techniques for self-localization, landmark-based approach is preferred for its simplicity and much less memory demanding for descriptions of robot surroundings. Door-plates are selected as visual landmarks. In this paper, we present an adaptive segmentation approach based on Principal Component Analysis (PCA) and scale-space filtering. To speed up the entire color segmentation and use the color information as a whole, PCA is implemented to project tristimulus R, G and B color space to the first principal component (1st PC) axis direction and scale-space filtering is used to get the centers of color classes. This method has been tested in the color segmentation of door-plate images captured by mobile robot CASIA-1. Experimental results are provided to demonstrate the effectiveness of this proposed method.

In this paper, a support vector machine (SVM) approach for automatic impulsive noise detection in corrupted image is proposed. Once the noises are detected, a filtering action based on regularization can be taken to restore the image. Experimental results show that the proposed SVM-based approach provides excellent performance with respect to various percentages of impulse noise.

The resource of image noise is analysized in this paper. Considering the image fuzzy generated in the process of image denoising in spatial field, the image denoising method based on wavelet neural network with orthonormal bases is elaborated. The denoising principle and construction method of orthonormal wavelet network is described. In the simulation experiment, median filtering, adaptive median filtering and sym wavelet neural network with orthonormal bases were used separately in the denoising for contaminated images. The experiment shows that, compared with traditional denoising method, image denoising method based on orthonormal wavelet neural network improves greatly the image quality and decreases the image ambiguity.

In order to enhance the contrast of an image, histogram equalization is wildly used. With global histogram equalization (GHE), the image is enhanced as a whole, and this may induce some areas to be overenhanced or blurred. Although local histogram equalization (LHE) acts adaptively to overcome this problem, it brings noise and artifacts to image. In this paper, a region-based enhancement algorithm is proposed, in which Grossberg network is employed to generate histogram and extract regions. Simulation results show that the image is obviously improved with the advantage of both GHE and LHE.

After performing discrete stationary wavelet transform (DSWT) to an image, local contrast is enhanced with non-linear operator in the high frequency sub-bands, which are at coarser resolution levels. In order to enhance global contrast for an infrared image, low frequency sub-band image is also enhanced employing non-incomplete Beta transform (IBT), simulated annealing algorithm (SA) and wavelet neural network (WNN). IBT is used to obtain non-linear gray transform curve. Transform parameters are determined by SA so as to obtain optimal non-linear gray transform parameters. Contrast type of original image is determined by a new criterion. Gray transform parameters space is determined respectively according to different contrast types. A kind of WNN is proposed to approximate the IBT in the whole low frequency sub-band image. The quality of enhanced image is evaluated by a total cost criterion. Experimental results show that the new algorithm can improve greatly the global and local contrast for images.

Using the Group Method of Data Handling (GMDH) a polynomial network is designed in this paper for learning the nonlinear image distortion of a camera. The GMDH network designed can effectively learn image distortion in various camera systems of different optical features unlike most existing techniques that assume a physical model explicitly. Compared to multilayer perceptrons (MLPs), which are popularly used to learn a nonlinear relation without modeling, a GMDH network is self-organizing and its learning is faster. We prove the advantages of the proposed technique with various simulated data sets and in a real experiment.

This paper designs an edge preserving regularization model for image restoration. First, we propose a generalized form of Digitized Total Variation (DTV), and then introduce it into restoration model as the regularization term. To minimize the proposed model, we map digital image onto network, and then develop energy descending schemes based on Hopfield neural network. Experiments show that our model can significantly better preserve the edges of image compared with the commonly used Laplacian regularization (with constant and adaptive coefficient). We also study the effects of neighborhood and gaussian parameter on the proposed model through experiments.

With a view to solve the problems in modern information science, we put forward a new subject named High-Dimensional Space Geometrical Informatics (HDSGI). It builds a bridge between information science and point distribution analysis in high-dimensional space. A good many experimental results certified the correctness and availability of the theory of HDSGI. The proposed method for image restoration is an instance of its application in signal processing. Using an iterative “further blurring-debluring-further blurring” algorithm, the deblured image could be obtained.

Although the baseline fractal image encoding algorithm could obtain very high compression ratio in contrast with other compression methods, it needs a great deal of encoding time, which limits it to widely practical applications. In recent years, an accelerating algorithm based on variance is addressed and has shortened the encoding time greatly; however, in the meantime, the image fidelity is obviously diminished. In this paper, a neural network is utilized to modify the variance-based encoding algorithm, which makes the quality of reconstructed images improved remarkably as the encoding time is significantly reduced. Experimental results show that the reconstructed images quality measured by peak-signal-to-noise-ratio is better than conventional variance-based algorithm, while the time consumption for encoding and the compression ratio are almost the same as the conventional variance-based algorithm.

We consider a class of auto-associative memories, namely, “associative cubes” in which gray-level images and the hidden orthogonal basis functions such as Walsh-Hadamard or Fourier kernels, are mixed and updated in the weight cubes,

C

. First, we develop an unsupervised learning procedure based upon the adaptive recursive algorithm. Here, each 2D training image is mapped into the associated 1D wavelet in the least-squares sense during the training phase. Second, we show how the recall procedure minimizes the recognition errors with a competitive network in the hidden layer. As the images corrupted by noises are applied to an associative cube, the nearest one among the original training images would be retrieved in the sense of the minimum Euclidean squared norm during the recall phase. The simulation results confirm the robustness of associative cubes even if the test data are heavily distorted by noises.

It has been shown that support vector machines (SVM) can be used in content-based image retrieval. Existing SVM based methods only extract low-level global or region-based features to form feature vectors and use traditional non-structured kernel function. However, these methods rarely consider the image structure or some new structured kernel types. In order to bridge the semantic gap between low-level features and high-level concepts, in this paper, a novel graph kernel based SVM method is proposed, which takes into account both low-level features and structural information of the image. Firstly, according to human selective visual attention model, for a given image, salient regions are extracted and the concept of Salient Region Adjacency Graph (SRAG) is proposed to represent the image semantics. Secondly, based on the SRAG, a novel graph kernel based SVM is constructed for image semantic retrieval. Experiments show that the proposed method shows better performance in image semantic retrieval than traditional method.

A modification of the well-known PicSOM retrieval system is presented. The algorithm is based on a variant of the self-organizing map algorithm that uses bootstrapping. In bootstrapping the feature space is randomly sampled and a series of subsets are created that are used during the training phase of the SOM algorithm. Afterwards, the resulting SOM networks are merged into one single network which is the final map of the training process. The experimental results have showed that the proposed system yields higher recall-precision rates over the PicSOM architecture.

Neural network is an important technique in many image understanding areas. Then the performance of neural network depends on the separative degree among the input vector extracted from an original image. However, most methods are not enough to understand the contents of a image. Accordingly, we propose a efficient method of extracting a spatial feature from a real image, and segmenting the TROI (: Textural Region Of Interest) from the clustered image without a pre-knowledge. Our approach presents the 2-passing k-means algorithm for extracting a spatial feature from image, and uses the unsupervised learning scheme for the block-based image clustering. Also, a segmentation of the clustered TROI is achieved by tuning 2D Gabor filter to the spatial frequency the clustered region. In order to evaluate the performance of the proposed method, the segmenting quality was measured according to the goodness based on the segmented shape of region. Our experimental results showed that the performance of the proposed method is very successful.

By right of the great convenience of computer graphics and digital imaging, it is much easier to alter the content of an image than before without any visually traces. Human has not believed what they see. Many digital images can not be judged whether they are real or feigned visually, i.e., many fake images are produced whose content is feigned. In this paper, firstly, image fakery is introduced, including how to produce fake images and its characters. Then, a fake image detection scheme is proposed, which uses radial basis function (RBF) neural network as a detector to make a binary decision on whether an image is fake or real. The experimental results also demonstrated the effectiveness of the proposed scheme.

A new approach to object detection using image icons based on Unit-linking PCNN (Pulse Coupled Neural Network) is introduced in this paper. Unit-linking PCNN, which has been developed from PCNN exhibiting synchronous pulse bursts in cat and monkey visual cortexes, is a kind of time-space-coding SNN (Spiking Neural Network). We have used Unit-linking PCNN to produce the global image icons with translation and rotation invariance. Unit-linking PCNN image icon (namely global image icons) is the 1-dimentional time series, and is a kind of image feature extracted from the time information that Unit-linking PCNN code the 2-dimentional image into. Its translation and rotation invariance is a good property in object detection. In addition to translation, rotation invariance, the object detection approach in this paper is also independent of scale variation.

In recent years digital watermarking was developed significantly and applied broadly for copyright protection and authentication. In this paper, a digital image watermarking scheme is developed using neural network to embedded watermark into DCT domain of each subimage blocks obtained by subsampling, which achieves adaptively watermark embedding and stronger robustness. Furthermore, in order to improve the security of the proposed watermarking, a random permutation process is used in watermarking process. Experiments show that the proposed watermarking scheme is effect and encouraging.

A simple and efficient inpaiting algorithm is proposed based on radial basis function network in this paper. Using the user defined areas, a neighborhood narrow band of the needing fixed pixels are computed by an erosion operator of mathematical morphology technique. Then the weights of RBF network are estimated and a continuous function is constructed, from which the needy inpainted pixels can be filled in.

No-reference perceptual quality assessment for JPEG images in real time is a critical requirement for some applications, such as in-service visual quality monitoring, where original information can not be available. This paper proposes a no-reference perceptual quality-assessment method based on a general regression neural network (GRNN). The three visual features of artifacts introduced in JPEG images are formulated block by block individually so that our method is computation-efficient and memory-efficient. The GRNN is used to realize the mapping of these visual features into a quality score in real time because of its excellent approximation and very short training time (one-pass learning). Experimental results on an on-line database show that our estimated scores have an excellent correlation with subjective MOS scores.

This paper provides the construction of statistical shape model based on elliptic Fourier transformation and minimum description length (MDL). The method does not require manual identification of landmarks on training shapes. Each training shapes can be decomposed into a set of ellipse by elliptic Fourier transformation at a different frequency level. The MDL objective function is based on elliptic Fourier descriptors and principal component analysis (EF-PCA). Experiments show that our method can get better models.

This paper presents a novel texture segmentation method using neural networks and a Markov random field (MRF) model. Multi-scale wavelet coefficients are used as input for the neural networks. The output of the neural network is modeled as

a posterior

probability. Initially, the multi-scale texture segmentation is performed by the posterior probabilities from the neural networks and MAP (maximum

a posterior

) classification. Then the MAP segmentation maps are produced at all scales. In order to obtain the more improved segmentation result at the finest scale, our proposed method fuses the multi-scale MAP segmentations sequentially from coarse to fine scales. This is done by computing the MAP segmentation given the segmentation map at one scale and

a priori

knowledge regarding contextual information which is extracted from the adjacent coarser scale segmentation. In this fusion process, the MRF prior distribution and Gibbs sampler are used, where the MRF model serves as the smoothness constraint and the Gibbs sampler acts as the MAP classifier.

This paper presents a likelihood estimation method from SOM (self organizing feature map), and texture segmentation is performed by using Bayesian estimation and SOM. Multi-scale wavelet coefficients are used as input for SOM, and likelihood probabilities for observations are obtained from trained SOMs. Texture segmentation is performed by the likelihood probability from trained SOMs and ML (maximum likelihood) classification. The result of texture segmentation is improved using contextual information. The proposed segmentation method performed better than segmentation method using HMT (hidden Markov trees) model. In addition, texture segmentation results by SOM and multi-scale Bayesian image segmentation technique called HMTseg also performed better than those by HMT and HMTseg.

In this paper, we proposed the image processing techniques for extracting the cracks in a concrete surface crack image and the ART1-based RBF network for recognizing the directions of the extracted cracks. The image processing techniques used are the closing operation of morphological techniques, the Sobel masking used to extract edges of the cracks, and the iterated binarization for acquiring the binarized image from the crack image. The cracks are extracted from the concrete surface image after applying two times of noise reduction to the binarized image. We proposed the method for automatically recognizing the directions (horizontal, vertical, -45 degree, 45 direction degree) of the cracks with the ART1-based network. The proposed ART1-based RBF network applied ART1 to the learning between the input layer and the middle layer and the Delta learning method to the learning between the middle layer and the output layer. The experiments using real concrete crack images showed that the cracks in the concrete crack images were effectively extracted and the proposed ART1-based RBF network was effective in the recognition of the direction of extracted cracks.

Detecting the presence of speech in a noisy signal is an unsolved problem affecting numerous speech processing applications. This paper shows an effective method employing support vector machines (SVM) for voice activity detection (VAD) in noisy environments. The use of kernels in SVM enables to map the data into some other dot product space (called feature space) via a nonlinear transformation. The feature vector includes the subband signal-to-noise ratios of the input speech and a radial basis function (RBF) kernel is used as SVM model. It is shown the ability of the proposed method to learn how the signal is masked by the acoustic noise and to define an effective non-linear decision rule. The proposed approach shows clear improvements over standardized VADs for discontinuous speech transmission and distributed speech recognition, and other recently reported VADs.

A new voice activity detection (VAD) method for microphone array signals is developed in this paper. A relatively pure speech signal can be obtained by applying noise canceling algorithms on some signals from microphone array. For suppressing correlated and uncorrelated noises, the proposed method doesn’t perform the same processing, but analyze the natures of the background noises by calculating the correlation between the noisy signals during silence intervals firstly. If the additive noises are correlated, relatively pure speech component is separated by blind source separation (BSS) method. Otherwise, this speech component is estimated by beamforming and maximum a posterior (MAP) algorithm. Then, a voice activity detection method based on entropy is employed to determine whether this relatively pure speech signal is active or not. Finally, this VAD result is used as reference to produce those of all array signals. Simulation results illustrate the validity of the proposed method.

In this paper, we propose a flexible two-stage algorithm for extracting desired periodic signals. In the first stage, if the period and phase information of the desired signal is available (or can be estimated), a minimum mean square error approach is used to coarsely recover the desired source signal. If only the period information is available (or can be estimated), a robust correlation based method is proposed to achieve the same goal. The second stage uses a higher-order statistics based Newton-like algorithm, derived from a constrained maximum likelihood criteria, to process the extracted noisy signal as cleanly as possible. A parameterized nonlinearity is adopted in this stage, adapted according to the estimated statistics of the desired signal. Compared with many existing extraction algorithms, the proposed algorithm has better performance, which is confirmed by simulations.

A principal component analysis (PCA) neural network (NN) based on signal eigen-analysis is proposed to blind signature waveform estimation in low signal to noise ratios (SNR) direct sequence synchronous code-division multiple-access (S-CDMA) signals. The received signal is firstly sampled and divided into non-overlapping signal vectors according to a temporal window, which duration is a period of signature waveform. Then an autocorrelation matrix is computed and accumulated by these signal vectors one by one. Since we have assumed that the synchronous point between the symbol waveform and observation window have been known, the signal vectors may be sampled and divided at the beginning of this synchronous point, therefore, each vector must contain all information of signature waveforms. In the end, the signature waveforms can be estimated by the principal eigenvectors of autocorrelation matrix blindly. Additionally, the eigen-analysis method becomes inefficiency when the estimated vector becomes longer. In this case, we can use the PCA NN method to realize the blind signature waveform estimation from low SNR input signals effectively.

A novel method for kernel design of a quadratic time frequency distribution (TFD) as the initial step for neural source estimation is proposed. The kernel is constructed based on the product ambiguity function (AF), which efficiently suppresses cross terms and noise in the ambiguity domain. In order to reduce the influence from the strong signal to the weak signal, an iterative approach is implemented. Simulation results validate the method and demonstrate suppression of cross terms and noise, and high resolution in the time frequency domain.

We designed a new channel estimator including two parts of neural network to estimate the amplitude and the angle of the frequency domain channel coefficients, respectively. The least mean error (LSE) is used for training. This neural network channel estimator (NNCE) makes full use of the learning property of the neural network (NN). Once the NN was trained, it reflected the channel fading trait of the amplitude and the angle respectively. It was no need of any matrix computation and it can get any required accuracy. It has been validated that the estimator is available in the pilot-symbol-aided (PSA) OFDM system.

In this paper, independent component analysis (ICA) is applied for feature extraction of non-Gaussian acoustic signals. The generalized Gaussian model (GGM) is introduced as the p.d.f. estimator in ICA because it can provide a general method for modeling non-Gaussian statistical structure of univariate distributions. It is demonstrated that the proposed method can efficiently extract ICA features for not only sup-Gaussian but also sub-Gaussian signals. The basis vectors are localized in both time and frequency domain and the resulting coefficients are statistically independent and sparse. The experiments of Chinese speech and the underwater signals show that the proposed method is more efficient than conventional methods.

Eye movements, cardiac signals, muscle noise and line noise,

etc

. present serious problems for the accuracy of Electroencephalographic (EEG) analysis. Some research results have shown that independent component analysis (ICA) can separate artifacts from multichannel EEG data. Further, considering the nonlinear dynamic properties of EEG signals, exponential analysis can be used to identify various artifacts and basic rhythms, such as

α

rhythm,

etc.

, from each independent component (IC). In this paper, we propose an automatic artifacts removal scheme for EEG data by combining ICA and exponential analysis. In addition, the proposed scheme can also be used to detect basic rhythms from EEG data. The experimental results on both the simulated data and the real EEG data demonstrate that the proposed scheme for artifacts removal has excellent performance.

Electromagnetic noise, unbalanced rotor noise and injuring bearing noise are three types of noise in faulting electromotor. An adaptive wavelet neural network is proposed to identify these noises. The process of wavelet-based feature extraction of signal is integrated into one part of neural network. During network’s training course the wavelet’s scale and shift parameters can be adaptively adjusted to fit input signal so that signal’s feature could be extracted in maximum limit. The network’s second part then uses these feature information to realize the identification of noise signal. The identification result of three types of noise of electromotor demonstrates that this neural network can give accurate identification result with high probability.

In this paper, the topic of fractional order digital differentiators (FODD) is designed using neural networks approximation method. First, FODD amplitude response is given in the form of sum of exponential basis functions. Then, the exponential basis function neural network is used to approximate FODD amplitude response. Finally, some examples compared with others’ method are given to illustrate the advantages of this paper approach.

In this paper, a new predictive algorithm for multivariate chaotic time series is proposed. Considering the correlations among time series, multivariate time series instead of univariate ones are taken as the inputs of predictive model. The model is implemented by a radial basis function neural network. To determine the number of model inputs, C-C method is applied to construct the embedding of the chaotic time series by choosing delay time window. The annual river runoff and annual sunspots are used in the simulation, and the proposed method is proven effective and valid.

Time series analysis is an important and complex problem in machine learning. In this paper, least squares support vector machine (LS-SVM) combined with particle swarm optimization (PSO) is used to time series prediction. The LS-SVM can overcome some shortcoming in the multilayer perceptron (MLP) and the PSO is used to tune the LS-SVM parameters automatically. A benchmark problem, Hénon map time series, has been used as an example for demonstration. It is showed this approach can escape from the blindness of man-made choice of the LS-SVM parameters. It enhances the efficiency and the capability of prediction.

The well-known mixtures of experts(ME) model is usually trained by expectation maximization(EM) algorithm for maximum likelihood learning. However, we have to first determine the number of experts, which is often hardly known. Derived from regularization theory, a regularized minimum cross-entropy(RMCE) algorithm is proposed to train ME model, which can automatically make model selection. When time series is modeled by ME, it is demonstrated by some climate prediction experiments that RMCE algorithm outperforms EM algorithm. We also compare RMCE algorithm with other regression methods such as back-propagation(BP) algorithm and normalized radial basis function(NRBF) network, and find that RMCE algorithm still shows promising results.

In this paper, based on the Volterra expansion of nonlinear dynamical system functions and the deterministic and nonlinear characterization of chaotic time series, the discrete Volterra neural networks are proposed to make prediction of chaotic time series. The predictive model of chaotic time series is established with the discrete Volterra neural networks and the steps of the learning algorithm with discrete Volterra neural networks are expressed. The predictive model and the learning algorithm are more effective and reliable than the adaptive higher-order nonlinear FIR filter. The Experimental and simulating results show the discrete Volterra neural networks can be successfully used to predict chaotic time series.

A new pre-processing method for regression is developed. The core idea is using three rules to clarify regression raw data. The rules are realized through introducing a judge function on the regression datum whose value determines the importance of the datum. By applying the rules, a new pre-processing method for regression is developed. Performance of the new method on a series of simulations demonstrate that it not only significantly increases computational efficiency and robustness, but also preserves generalization capability of a regression method. Incorporated with any regression method, the developed method then can be efficiently applied to regression of large data sets.

An improved radial basis function neural network (IRBFNN) with unsymmetrical Gaussian function is presented to simplify the structure of RBFNN. The improved resource allocating network (IRAN) is developed to design IRBFNN online for nonlinear dynamic system modeling, integrating the typical resource allocating network (RAN) with merging method for similar hidden units, deleting strategy for redundant hidden units, and LMS learning algorithm with moving data window for output link weights of network. The proposed approach can effectively improve the precision and generalization of IRBFNN. The combination of IRBFNN and IRAN is competent for the online modeling of nonlinear dynamic systems. The feasibility and effectiveness of the modeling method have been demonstrated by simulations.

The least squares support vector machines (LS-SVMs) is a kernel method. The training problem of LS-SVMs is solved by finding a solution to a set of linear equations. This makes online adaptive implementation of the algorithm feasible. An improved adaptive algorithm is proposed for training the LS-SVMs in this paper. This algorithm is especially useful on online nonlinear system modeling. The experiments with benchmark problem have shown the validity of the proposed method even in the case of additive noise to the system.

A single neural network model developed from a limited amount of sample data usually lacks robustness and generalization. Neural network model robustness and prediction accuracy can be improved by combining multiple neural networks. In this paper a new method of the multiple neural networks for nonlinear modeling is proposed. A whole training sample data set is partitioned into several subsets with different centers using fuzzy c-means clustering algorithm (FCM), and the individual neural network is trained by each subset to construct the subnet respectively. The degrees of memberships are used for combining the outputs of subnets to obtain the final result, which are gained from the relationship between a new input sample data and each cluster center. This model has been evaluated and applied to estimate the status-of-loose of jig washer bed. Simulation results and actual application demonstrate that this model has better generalization, better prediction accuracy and wider potential application online.

A new reproducing kernel in reproducing kernel Hilbert space (RKHS), namely the multi-resolution reproducing kernel, is presented in this paper. The multi-resolution reproducing kernel is generated by scaling basis function at some scale and wavelet basis function with different resolution. Based on multi-resolution reproducing kernel and

ν

- support vector regression (

ν

-SVR) method, a new regression model is proposed. The regression model used to nonlinear system identification, incorporate the advantage of the support vector machines and the multi-resolution property of wavelet. Simulation examples are given to illustrate the feasibility and effectiveness of the method.

In this paper a new structure of a recurrent neurofuzzy network is proposed. The network considers two cascade-interconnected Fuzzy Inference Systems (FISs), one recurrent and one static, that model the behaviour of a unknown dynamic system from input-output data. Each FIS’s rule involves a linear system in a controllable canonical form. The training for the recurrent FIS is made by a gradient-based Real-Time Recurrent Learning Algorithm (RTRLA), while the training for the static FIS is based on a simple gradient method. The initial parameter conditions previous to training are obtained by extracting information from a static FISs trained with delayed input-output signals. To demonstrate its effectiveness, the identification of two non-linear dynamic systems is included.

This paper proposes a dynamic recurrent fuzzy wavelet network (RCFWN) for identified nonlinear dynamic systems. Temporary relations are embedded in the network by adding feedback connections in the second layer of the fuzzy wavelet network. In addition, the study algorithm of the RCFWN is introduced and its stability analysis is studied. Finally, the RCFWN is applied in several simulations. The results verify the effectiveness of the RCFWN.

This paper analyzes various formulations for the recursive training of neural networks that can be used for identifying and optimizing nonlinear processes on line. The study considers feedforward type networks (FFNN) adapted by three different methods: inverse Hessian matrix approximation, calculation of the inverse Hessian matrix using a Gauss-Newton recursive sequential algorithm, and calculation of the inverse Hessian matrix in a recursive type Gauss-Newton algorithm. The study is completed using two network structures that are linear in the parameters: a radial basis network and a principal components network, both trained using a recursive least squares algorithm. The corresponding algorithms and a comparative test consisting of the on-line estimation of a reaction rate are detailed. The results indicate that all the structures were capable of converging satisfactorily in a few iteration cycles, FFNN type networks showing better prediction capacity, but the computational effort of the recursive algorithms is greater.

We introduce a consecutive identification of ANFIS-based fuzzy systems with the aid of genetic data granulation to carry out the model identification of complex and nonlinear systems. The proposed model implements system structure and parameter identification with the aid of information granulation and genetic algorithms. The design methodology emerges as a hybrid structural optimization and parametric optimization. Information granulation realized with HCM clustering algorithm help determine the initial parameters of fuzzy model such as the initial apexes of the membership functions in the premise and the initial values of polynomial functions in the consequence. And the structure and the parameters of fuzzy model are identified by GAs and the membership parameters are tuned by GAs. In this case we exploit a consecutive identification. The numerical example is included to evaluate the performance of the proposed model.

In this study, we propose the consecutive optimization of ANFIS-based fuzzy systems with fuzzy set. The proposed model formed by using respective fuzzy spaces (fuzzy set) implements system structure and parameter identification with the aid of information granulation and genetic algorithms. Information granules are sought as associated collections of objects (data, in particular) drawn together by the criteria of proximity, similarity, or functionality. Information granulation realized with HCM clustering help determine the initial parameters of fuzzy model such as the initial apexes of the membership functions in the premise and the initial values of polynomial functions in the consequence. And the initial parameters are tuned with the aid of the genetic algorithms and the least square method. To optimally identify the structure and parameters we exploit the consecutive optimization of ANFIS-based fuzzy model by means of genetic algorithms. The proposed model is contrasted with the performance of conventional fuzzy models in the literature.

Based on the time evolution similarity principle of the topological neighbor phase points in the Phase Space Reconstruction (PSR), a new modeling approach of Short-Term Load Forecasting (STLF) with integrated dynamics mechanism and based on the fusion of the dynamic optimal neighbor phase points (DONP) and Improved Chaotic Neural Networks (ICNN) model was presented in this paper. The ICNN model can characterize complicated dynamics behavior. It possesses the sensitivity to the initial load value and to the walking of the whole chaotic track. The input dimension of ICNN is decided using PSRT, and the training samples are formed by means of the stepping dynamic space track on the basis of the DONP. So it can improve associative memory and generalization ability of ICNN model. The testing results show that proposed model and algorithm can enhance effectively the precision of STLF and its stability.

A stable adaptive neural network control approach is proposed in this paper for uncertain nonlinear strict-feedback systems based on backstepping. The key assumptions are that the neural network approximation errors satisfy certain bounding conditions. By a special scheme, the controller singularity problem is avoided perfectly. The proposed scheme improves the control performance of systems and extends the application scope of nonlinear systems. The overall neural network control systems guarantee that all the signals of the systems are uniformly ultimately bounded and the tracking error converges to a small neighborhood of zero by suitably choosing the design parameter.

In this paper, we address, in a backstepping way, stabilization problem for a class of switched nonlinear systems whose subsystem with trigonal structure by using neural network. An adaptive neural network switching control design is given. Backsteppping, domination and adaptive bounding design technique are combined to construct adaptive neural network stabilizer and switching law. Based on common Lyapunov function approach, the stabilization of the resulting closed-loop systems is proved.

A robust adaptive control scheme is presented for a class of uncertain continuous-time multi-input multi-output (MIMO) nonlinear systems. Within these schemes, multiple multi-layer neural networks are employed to approximate the uncertainties of the plant’s nonlinear functions and robustifying control term is used to compensate for approximation errors. All parameter adaptive laws and robustifying control term are derived based on Lyapunov stability analysis so that all the signals in the closed loop are guaranteed to be semi-globally uniformly ultimately bounded and the tracking error of the output is proven to converge to a small neighborhood of zero. While the relationships among the control parameters, adaptive gains and robust gains are established to guarantee the transient performance of the closed loop system.

This paper focuses on the design of passive controller with adaptive neural compensation for uncertain strict-feedback nonlinear systems with input-delay. For local linearization model, the delay-dependent

γ

-passive control is presented. Then,

γ

-passive control law of local linear model is decomposed as the virtual control of sub-systems by backstepping. In order to compensate the nonlinear dynamics, the adaptive neural model is proposed. The NN weights are turned on-line by Lyapunov stability theory with no prior training. The design procedure of whole systems is a combination of local

γ

-passive control and adaptive neural network compensation techniques.

Multiple radial based function (RBF)neural network models are used to cover the uncertainty of time variant nonlinear system, and multiple element controllers are set up based on the multiple RBF models. At every sample time, the closest model is selected by an index function which is formed by the integration of model output error. The element controller based on this model will be switched as the controller of the controlled system. This kind of multiple model adaptive controller (MMAC)is an extension of the MMAC in linear system, and it can improve the transient response and performance of the controlled system greatly.

Presented in this paper is neural network based adaptive control for a class of affine nonlinear systems in the strict-feedback form with unknown nonlinearities. A popular recursive design methodology – backstepping is employed to systematically construct feedback control laws and associated Lyapunov functions. The significance of this paper is to make best use of available signals, avoid unnecessary parameterization, and minimize the node number of neural networks as on-line approximators. The design assures that all the signals in the closed loop are semi-globally uniformly, ultimately bounded and the outputs of the system converges to a tunable small neighborhood of the desired trajectory. Novel parameter tuning algorithms are obtained on a more practical basis.

A new control scheme using multiple models and RBF neural networks is developed in this paper. The proposed scheme consists of multiple feedback linearization controllers, which are based on the known nominal dynamics model and a compensating controller, which is based on RBF neural networks. The compensating controller is applied to improve the transient performance. The neural network is trained online based on Lyapunov theory and learning convergence is thus guaranteed. Simulation results are presented to demonstrate the validity of the proposed method.

A novel robust adaptive neural network control (RANNC) is proposed for a class of strict-feedback nonlinear systems with both unknown system nonlinearities and unknown virtual control gain nonlinearities. The synthesis of RANNC is developed by use of the input-to-state stability (ISS), the backstepping technique, and generalized small gain approach. The key feature of RANNC is that the order of its dynamic compensator is only identical to the order

n

of controlled system, such that it can reduce the computation load and makes particularly suitable for parallel processing. In addition, the possible controller singularity problem can be removed elegantly. Finally, simulation results are presented to validate the effectiveness of the RANNC algorithm.

A neural network based robust adaptive control design scheme is developed for a class of nonlinear systems represented by input-output models with an unknown nonlinear function and unmodeled dynamics. By on-line approximating the unknown nonlinear functions and unmodeled dynamics by radial basis function (RBF) networks, the proposed approach does not require the unknown parameters to satisfy the linear dependence condition. It is proved that with the proposed control law, the closed-loop system is stable and the tracking error converges to zero in the presence of unmodeled dynamics and unknown nonlinearity.

A neural-network-based robust output feedback H

 ∞ 

control design is suggested for control of a class of nonlinear systems with time delays. The design approach employs a neural network, of which the activation functions satisfy the sector conditions, to approximate the delayed nonlinear system. A full-order dynamic output feedback controller is designed for the approximating neural network. The closed-loop neural control system is transformed into a novel neural network model termed standard neural network model (SNNM). Based on the robust H

 ∞ 

performance analysis of the SNNM, the parameters of output feedback controllers can be obtained by solving some lilinear matrix inequalities (LMIs). The well-designed controller ensures the asymptotic stability of the closed-loop system and guarantees an optimal H

 ∞ 

norm bound constraint on disturbance attenuation for all admissible uncertainties.

In this paper, a support vector machine (SVM) with polynomial kernel function enhanced nonlinear self-tuning controller is developed, which combines the SVM identifier and parameters’ modifier together. The inverse model of a nonlinear system is achieved by off-line black-box identification according to input and output data. Then parameters of the model are modified online using gradient descent algorithm. Simulation results show that SVM based self-tuning control can be well applied to nonlinear uncertain system. And the SVM based self-tuning control of nonlinear system has good robustness performance in tracking reference input with good generalization ability.

In this paper, a design procedure of support vector machine (SVM) with RBF kernel function based internal model control (IMC) strategy for stable nonlinear systems with input-output form is proposed. The control scheme consists of two controllers: a SVM based controller which fulfils the direct inverse model control and a traditional controller which fulfils the close-loop control. And so the scheme can deal with the errors between the process and the SVM based internal model generated by model mismatch and additional disturbance. Simulations are given to illustrate the proposed design procedure and the properties of the SVM based internal model control scheme for unknown nonlinear systems with time delay.

Based on fast online support vector regression (SVR) algorithm, reverse model of system model is constructed, and adaptive internal model controller is developed. First, SVR model and its online training algorithm are introduced. A kernel cache method is used to accelerate the online training algorithm, which makes it suitable for real-time control application. Then it is used in internal model control (IMC) for online constructing internal model and designing the internal model controller. Output errors of the system are used to control online SVR algorithm, which made the whole control system a closed-loop one. Last, the fast online SVM based adaptive internal model control was used to control a benchmark nonlinear system. Simulation results show that the controller has simple structure, good control performance and robustness.

A variable structure control (VSC) scheme for linear black-box multi-input/multi-output (MIMO) systems based on support vector machine (SVM) is developed. After analyzing character of MIMO system, an additional control is designed to track trajectory. Then VSC algorithm is adopted to eliminate the difference. By estimating outputs of next step, VSC inputs and additional inputs are obtained directly by two kinds of trained SVMs, and so recognition of system parameters is avoided. A linear MIMO system is introduced to prove the scheme, and simulation shows that the high identification precision and quick training speed.

A novel least squares support vector machines based on Mexican hat wavelet kernel is presented in the paper. The wavelet kernel which is admissible support vector kernel is characterized by its local analysis and approximate orthogonality, and we can well obtain estimates for regression by applying a least squares wavelet support vector machines (LS-WSVM). To test the validity of the proposed method, this paper demonstrates that LS-WSVM can be used effectively for the identification and adaptive control of nonlinear dynamical systems. Simulation results reveal that the identification and adaptive control schemes suggested based on LS-WSVM gives considerably better performance and show faster and stable learning in comparison to neural networks or fuzzy logic systems. LS-WSVM provides an attractive approach to study the properties of complex nonlinear system modeling and adaptive control.

Combining multiple neural networks appears to be a very promising approach for improving neural network generalization since it is very difficult, if not impossible, to develop a perfect single neural network. Therefore in this paper, a nonlinear model predictive control (NMPC) strategy using multiple neural networks is proposed. Instead of using a single neural network as a model, multiple neural networks are developed and combined to model the nonlinear process and then used in NMPC. The proposed technique is applied to water level control in a conic water tank. Application results demonstrate that the proposed technique can significantly improve both setpoint tracking and disturbance rejection performance.

This paper considers the problem of stabilizing a black-box plant with time delay using an improved double controller scheme. The PID parameters of the load controller of the double-controller scheme are obtained by a neural network controller with back propagation algorithm. Based on the adaptive algorithm of Universal Learning Network (ULN), ULN is adopted for modeling the plant and being a predictor of the control system. Simulation results prove the applicability and effectiveness of the improved double-controller scheme. ULN and the neural network controller give the double-controller scheme more representing abilities and robust ability.

In this paper, we present a new sliding mode controller for a class of unknown nonlinear discrete-time systems. We make the following two modifications: 1) The neural identifier which is used to estimate the unknown nonlinear system, applies new learning algorithms. The stability and non-zero properties are proved by dead-zone and projection technique. 2) We propose a new sliding surface and give a necessary condition to assure exponential decrease of the sliding surface. The time-varying gain in the sliding mode produces a low-chattering control signal. The closed-loop system with sliding mode controller and neural identifier is proved to be stable by Lyapunov method.

This paper addresses a new type of control framework for dynamical stochastic systems, which is called

statistic tracking control

here. General non-Gaussian systems are considered and the tracked objective is the statistic information (including the moments and the entropy) of a given target probability density function (PDF), rather than a deterministic signal. The control is aiming at making the statistic information of the output PDFs to follow those of a target PDF. The B-spline neural network with modelling error is applied to approximate the corresponding dynamic functional. For the nonlinear weighting system with time delays in the presence of exogenous disturbances, the generalized H

2

and H

 ∞ 

optimization technique is then used to guarantee the tracking, robustness and transient performance simultaneously in terms of LMI formulations.

The main idea of this paper is to characterize the uncertainties of control system base upon entropy concept. The wavelet neural networks is used to approach the nonlinear system through minimizing Renyi’s entropy criterion of the system estimated error, and the controller design is based upon minimizing Renyi’s entropy criterion of the system tracking errors. An illustrative example is utilized to demonstrate the effectiveness of this control solution, and satisfactory results have been obtained.

For a class of nonlinear stochastic Markovian jump systems, a novel feedback control law design is presented, which includes three steps. Firstly, the multi-layer neural networks are used to approximate the nonlinearities in the different jump modes. Secondly, the overall system is represented by the mode-dependent linear difference inclusion, which is suitable for control synthesis based on Lyapunov stability. Finally, by introducing stochastic quadratic performance cost, the existence of feedback control law is transformed into the solvability of a set of linear matrix inequalities. And the optimal upper bound of stochastic cost can be efficiently searched by means of convex optimization with global convergence assured.

Biologically inspired soft computing paradigms such as neural networks are popular learning models adopted in adaptive control systems for their ability to cope with a changing environment. However, continual changes induce uncertainty that limits the applicability of conventional validation techniques to assure a reliable system performance.

In this paper, we present a dynamic approach to estimate the performance of two types of neural networks employed in an adaptive flight controller: the validity index for the outputs of a Dynamic Cell Structure (DCS) network and confidence levels for the outputs of a Sigma-Pi (or MLP) network. Both tools provide statistical inference of the neural network predictions and an estimate of the current performance of the network. We further evaluate how the quality of each parameter of the network (e.g., weight) influences the output of the network by defining a metric for parameter sensitivity and parameter confidence for DCS and Sigma-Pi networks. Experimental results on the NASA F-15 flight control system demonstrate that our techniques effectively evaluate the network performance and provide validation inferences in a real-time manner.

The design of self-organizing fuzzy control (SOFC) systems does not rely on experts’ knowledge. The systems can establish fuzzy reasoning rules and adjust fuzzy parameters on-line automatically. So they are suitable for the control of plants for which we have no appropriate mathematical models and no other knowledge. There are three kinds of SOFCs, i.e., conventional, neural networks based and genetic algorithm based ones. New achievements of the above SOFC are summarized in this paper.

This paper presents the application of an evolutionary neural network controller in a stabilisation problem involving an inverted pendulum. It is guaranteed that the resulting continuous closed-loop system is asymptotically stable. The process of training the neural network controller can be treated as a constrained optimisation problem where the equality constraint is derived from the Lyapunov stability criteria. The decision variables in this investigation are made up from the connection weights in the neural network, a positive definite matrix required for the Lyapunov function and a matrix for the stability constraint while the objective value is calculated from the closed-loop system performance. The optimisation technique chosen for the task is a variant of genetic algorithms called a cooperative coevolutionary genetic algorithm (CCGA). Two control strategies are explored: model-reference control and optimal control. In the model-reference control, the simulation results indicate that the tracking performance of the system stabilised by the evolutionary neural network is superior to that controlled by a neural network, which is trained via a neural network emulator. In addition, the system stabilised by the evolutionary neural network requires the energy in the level which is comparable to that found in the system that uses a linear quadratic regulator in optimal control. This confirms the usefulness of the CCGA in nonlinear system stabilisation applications.

A decentralized neural network controller for a class of large-scale nonlinear systems with the higher-order interconnections is proposed. The neural networks (NNs) are used to cancel the effects of unknown subsystems, while the robustifying terms are used to counter the effects of the interconnections. Semi-global asymptotic stability results are obtained and the tracking error converges to zero.

A pseudo-linear radial basis function (PLRBF) network is developed in this paper. This network is used to model a real process and its weights are on-line updated using a recursive orthogonal least squares (ROLS) algorithm. The developed adaptive model is then used in model predictive control strategy, which is applied to a pilot multivariable chemical reactor. The first stage of the project, simulation study, has been investigated and is presented. The effectiveness of the adaptive control in improving the closed-loop performance has been demonstrated for process time-varying dynamics and model-process mismatch.

To improve BP algorithm in overcoming the local minimum problem and accelerating the convergence speed, a new improved algorithm based on global revision factor of BP neural network is presented in this paper. The basic principle is to improve the formula of weight adjusting used momentum back propagation. A global revision factor is added in the weight value adjusting formula of momentum BP. Faster learning speed is obtained by adaptive adjusting this factor. The new BP algorithm is compared with other improved BP algorithms on many aspects. Simulation and applications for complex nonlinear function approximation, neural PID parameter tuning indicates that it has better training speed and precision than momentum back propagation and adaptive learning rate.

Reliable control of superheated steam temperature is necessary to ensure high efficiency and high load-following capability in the operation of modern power plant. A nonlinear generalized predictive controller based on neuro-fuzzy network (NFGPC) is proposed in this paper, which consists of local GPCs designed using the local linear models of the neuro-fuzzy network. The proposed nonlinear controller is applied to control the superheated steam temperature of a 200MW power plant, in which much better performance than the traditional cascade PI controller or the linear GPC is obtained.

In this paper, a feedforward active noise control (ANC) system using a mode-free neural network (MFNN) controller based on simultaneous perturbation stochastic approximation (SPSA) algorithm is considered. The SPSA-based MFNN control algorithm employed in the ANC system is first derived. Following this, computer simulations are carried out to verify that the SPSA-based MFNN control algorithm is effective for a nonlinear ANC system. Simulation results show that the proposed scheme is able to significantly reduce disturbances without the need to model the secondary-path and has better tracking ability under variable secondary-path. This observation implies that the SPSA-based MFNN controller frees the ANC system from the modeling of the secondary-path.

The AC-excited hydrogenerator (ACEH) is a novel type of hydraulic generation system. Concern about its integrative control strategy is increasing, owing to the features of uncertain and nonlinear as well as parameters coupling and time-variation for three parts of water flux, hydroturbine and generator. A cascade-connected self-adaptive fuzzy-neural network control strategy is proposed, which the former controller uses a self-tuning fuzzy algorithm with the intelligent weight function rulers, the latter adopts a self-adaptive neural network controller based on dynamical coupling characteristics of controlled plants. By comparison with traditional PID control, simulation results have shown that this hydrogenerator system appears good robustness against load disturbance and system parameters uncertainty.

Since Magneto-rheological (MR) suspension has nonlinearity and time-delay, the application of linear feedback strategy has been limited. This paper addresses the problem of control of MR suspension with time-delay when transient dynamics are presented. An adaptive Fuzzy-Neural Network Control (FNNC) scheme for the transient course is proposed using fuzzy logic control and artificial neural network methodologies. To attenuate the adverse effects of time-delay on control performance, a Time Delay Compensator (TDC) is established. Then, through a numerical example of a quarter car model and a real road test with a bump input, the comparison is made between passive suspension and semi-active suspension. The results show that the MR vehicle with FNNC strategy can depress the peak acceleration and shorten the setting time, and the effect of time-delay can be attenuated. The results of road test with the similarity of numerical study verify the feasibility of the control strategy.

An approach base on an adaptive fuzzy basis function network (AFBFN) is presented for fault-tolerance treatment in uncertain power systems. The uncertain system is composed of unknown part and known part represented by a mathematical model. A fuzzy basis function network (FBFN) is trained offline to represent the model of unknown part. AFBFN is trained online to represent the unknown model included the unknown fault. The reference model is composed of the known mathematical model and FBFN. According to outputs of actual system, AFBFN and reference model, another AFBFN is adopted to complete the fault-tolerance process and obtain the feedback control input of the uncertain system, which makes the actual system to track output of the reference model. A simulation example of the multi-machine coupling power systems is given to validate the method. The result proved its effectiveness.

Study of maintaining safe operation of marine main engine while one or several faults occur in marine systems has been put a high value recently in field of shipbuilding industry. The paper establishes a dynamic mathematical model of marine main engine system and analyses its characteristics while it is in abnormal conditions. Afterwards, a fault tolerant control system with artificial neural network algorithm is designed for improvement of operation of main engine that is in abnormal operation under consideration that a fault tolerant control system is able to fulfill function of fault toleration. The results of simulation experiments show that this fault tolerant control system is suitable for safe operation of marine main engine system.

Due to the nonlinear characteristics of Double Salient Permanent Magnet (DSPM) motor, the fixed-gain Proportional Integer (PI) controller can not perform well at all operating conditions. To increase the robustness of PI controllers, we present a self-tuning PI controller for speed control of DSPM motor drive system. The method is systematic and robust to parameter variations. We first treat the model of the DSPM motor drive. A well-trained multi-layer Neural Network (NN) is used to map the nonlinear relationship between the controller coefficients (Kp, Ki) and the control parameters (switching angles and current). Then we apply genetic algorithm to optimize the coefficients of the PI controller. A second NN is used to evaluate the fitness value of each chromosome in programming process of genetic algorithm. A main advantage of our method is that we do not require the accurate model of DSPM motor (which is always difficult to acquire), and the training process of NN can be done off-line through personnel computer, so that the controller can be implemented with a Digital Signal Processor (DSP-TMS320F2407). The experimental results illuminated that the proposed variable PI controller offers faster dynamic response and better adaptability over wider speed range.

This study introduces an efficient speed controller for a DC servomotor based on neural noncausal inverse modeling of the motor. For this mission; first, motor mathematical model is obtained in digital form. Secondly, to be able to generate necessary inputs which drive the plant, open loop control signals, the inverse model of the system is identified by an ANN structure. Then, a neural controller is introduced immediately, which is trained by a composite error signal. During the identification and control process, an efficient numerical computing based on Newton-Raphson method simulates the dynamic of the motor. The success of the designed control system is tested by a simulation study considering real conditions to be able to occur in real-time running of the system.

A novel dynamic time delay neural network is proposed for ultrasonic motors identification and control in this paper. By introducing time delay neurons, the neural network identifier and controller of ultrasonic motors are constructed. Both of them are trained online by using an improved back-propagation algorithm. The usefulness and validity of the presented algorithm is examined by the experiments.

A Gaussian basis function based CMAC (GCMAC) is proposed for the feed-forward control of line motor. It has both the advantages of CMAC and GBF (Gaussian basis function), such as lower memory requirement, faster converging speed and more accurate approximation. Considering that the GCMAC’s parameters selection is crucial for linear motor to get better control performance, we employ a particle swarm optimization algorithm to search for the optimal learning rate of the GCMAC. A numerical example of a linear motor model in wafer stage is preformed. The simulation results verify the effectiveness of the PSO-optimized GCMAC feed-forward controller.

The basic assumption that the angular velocity of the input crank is constant in much mechanism synthesis and analysis may not be validated when an electric motor is connected to driven then mechanism. First, the controller-motor-mechanism coupling system is studied in this paper, numerically simulation result demonstrate the crank angular speed fluctuations for the case of a constant voltage supply to DC motor. Then a novel algorithm of motor-mechanism adaptive PID control with BP neural network is proposed, using the approximate ability to any nonlinear function of the neural network. The neural network are used to predicted models of the controlled variable, this information is transferred to PID controller, through the readjustment of the pre-established set. The simulation results show that the crank speed fluctuation can be reduced substantially by using feedback control.

In this paper, the automatic balancing method is studied at constant speed or acceleration operation for rotor systems. The magnitude and phase of original imbalance is determined by the influence coefficient method through measurement of vibrations. In addition, a self-tuning neural-PD controller is designed to control the angular positions of correction masses for the automatic balancer. The equations of motion for a cylindrical rotor with radial imbalances are modeled to simulate two-plane balancing by means of axis-passed balancers. The dynamic responses before and after automatic balancing are investigated to justify the validity of the present method.

This paper considers the problem of PD control of overhead crane in the presence of uncertainty associated with crane dynamics. By using radial basis function neural networks, these uncertainties can be compensated effectively. This new neural control can resolve the two problems for overhead crane control: 1) decrease steady-state error of normal PD control. 2) guarantee stability via neural compensation. By Lyapunov method and input-to-state stability technique, we prove that these robust controllers with neural compensators are stable. Real-time experiments are presented to show the applicability of the approach presented in this paper.

A hybrid actuator is presented in this paper. Hybrid actuator is a new type of planar parallel robot, and requires precise control of the position of the mechanism. In order to achieve the desired accuracies, nonlinear factors as friction must be accurately compensated in the real-time servo control algorithm. According to the characteristics of the hybrid actuator, a hybrid intelligent control algorithm based on PID control and cerebellar model articulation control (CMAC) techniques was presented and used to perform control of hybrid actuator for the first time. Simulation results show that this method can improve the control effect remarkably compared with the tradi- tional control strategy.

In this paper, a fuzzy inference system based on support vector machi- nes is proposed for nonlinear system control. Support vector machines provides a mechanism to extract support vectors for generating fuzzy if-then rules from the training data set, and a method to describe the fuzzy inference system in terms of kernel functions. Thus it has the inherent advantages that the model doesn’t have to decide the number of fuzzy rules in advance, and has universal approximation ability and good generalization ability. The simulation results for stabilizing control of double inverted pendulum system are provided to show the validity and applicability of the proposed method.

Recently, based on multi-resolution analysis, wavelet neural networks (WNN) have been proposed as an alternative to NN for approximating arbitrary nonlinear functions in

L

2

(

R

). Discontinuous friction function is an unavoidable nonlinear effect that can limit control performance in mechanical systems. In this paper, adaptive WNN is used to design a friction compensator for a single joint mechanical system. Then asymptotically stability of the system is assured by adding a PD controller and adaptive robust terms. The simulation results show the validity of the control scheme.

An important application of mobile robots is searching a region to locate the origin of a specific phenomenon. A variety of optimization algo-rithms can be employed to locate the target source, which has the maximum intensity of the distribution of some detected function. We propose a neural network based dual heuristic programming (DHP) algorithm to solve the collective robotic search problem. Experiments were carried out to investigate the effect of noise and the number of robots on the task performance, as well as the expenses. The experimental results were compared with those of stochastic optimization algorithm. It showed that the performance of the dual heuristic programming (DHP) is better than the stochastic optimization method.

Hyper-redundant manipulator has more degrees of freedom than the least necessary to perform a given task, thus it has the features of overcoming conventional industrial robot’s limitation to carry out a designated difficult task. When the manipulator carries out the missions such as brushing or writing on a surface, drilling or inspection in a hole, the end-effector of the manipulator usually has both position and orientation requirement. Effective control of the hyper-redundant manipulator with such constrained end-effector is difficult for its redundancy. In this paper, a novel approach based on RBF neural network has been proposed to kinematically control the hyper-redundant manipulator. This technique, using variable regular polygon and RBF neural networks models, is completely capable of solving the control problem of a planar hyper-redundant manipulator with any number of links following any desired direction and path. With the shape transformation of variable regular polygon, the manipulator’s configuration changes accordingly and moves actively to perform the tasks. Compared with other methods to our knowledge, this technique has such superiorities as fewer control parameters and higher precision. Simulations have demonstrated that this control technique is available and effective.

A new robust adaptive neural networks tracking control with online learning controller is proposed for robot systems. A learning strategy and robust adaptive neural networks are combined into a hybrid robust control scheme. The proposed controller deals mainly with external disturbances and nonlinear uncertainty in motion control. A neural network (NN) is used to approximate the uncertainties in a robotic system. Then the disadvantageous effects on tracking performance, due to the approximating error of the NN in robotic system, are attenuated to a prescribed level by an adaptive robust controller. The learning techniques of NN will improve robustness with respect to uncertainty of system, as a result, improving the dynamic performance of robot system. A simulation example demonstrates the effectiveness of the proposed control strategy.

This paper addresses the problem of controlling an acrobot, an under-actuated robotic systems, using fuzzy neural network approach. A five-layer Takagi-Sugeno fuzzy neural network control (TSFNNC) is proposed to swing up the acrobot from the low stable equilibrium to approach and balance around its top unstable equilibrium position. By analyzing the system dynamics, total energy and potential energy of the system are introduced in the second layer, with the system states as the inputs to the first layer. Fuzzy membership functions and rules are depicted in the third and fourth layers respectively. The fifth layer works as the final output. A modified particle swarm optimizer (PSO) is adopted to train the consequents in the fourth layer. Simulation results indicate that the integrated TSFNNC approach can control the acrobot system from upswing to balance process effectively. This approach provides an easy and feasible solution for similar control problems.

A kind of recurrent fuzzy wavelet neural network (RFWNN) is constructed by using recurrent wavelet neural network (RWNN) to realize fuzzy inference. In the network, temporal relations are embedded in the network by adding feedback connections on the first layer of the network, and wavelet basis function is used as fuzzy membership function. An adaptive control scheme based on RFWNN is proposed, in which, two RFWNNs are used to identify and control plant respectively. Simulation experiments are made by applying proposed adaptive control scheme on robotic tracking control problem to confirm its effectiveness.

A novel dynamic control scheme for nonholonomic mobile robots is developed in this paper. The dynamics of mobile robot based on improved radial basis function neural network (IRBFNN) is modeled online by the improved algorithm of resource allocating network (IRAN). The control scheme of mobile robot integrates a velocity controller based on backstepping technology and a torque controller based on the IRBFNN and robust compen-sator. The simulations have shown that the control system is competent for the robust tracking control of mobile robot.

In this paper, we propose a grasping control method for robot hand using fuzzy theory and partially- linearized neural network. The robot hand has Double-Octagon Tactile Sensor (D.O.T.S), which has been proposed in our previous papers, to detect grasping force between the grasped object and the robot fingers. Because the measured forces are fluctuant due to the measuring error and vibration of the hand, the tactile information is ambiguous. In order to quickly control the grasping force to prevent the grasped object sliding out off the robot fingers, we apply the possibility theory to deal with the ambiguous problem of the tactile information, and use the partially- linearized neural network (P.L.N.N) to construct a fuzzy neural network. The method proposed in this paper is verified by applying it to practical grasping control of breakable objects, such as eggs, fruits, etc.

A position neurocontroller for robot manipulators with a tendon-driven transmission system has been developed allowing to track desired trajectories and reject external disturbances. The main problem to control tendons proceeds from the different dimensions between the joint and the tendon spaces. In order to solve this problem we propose a static neural network in cascade with a torque resolutor. The position controller is built as a parametric neural network by using basis functions obtained directly from the finger structure. This controller insure that the tracking error converges to zero and the weights of the network are bounded. The implementation has been improved partitioning the neural network into subnets and using the Kronecker product. Both control and weight updating laws have been designed by means of a Lyapunov energy function. In order to improve the computational efficient of the neural network, this has been split up into subnets to compensate inertial, Coriolis/centrifugal and gravitational effects. The NN weights are initialised at zero and tuned on-line with no ”off-line learning phase”. This scheme has been applied to an anthropomorphic robotic finger with a transmission system based on tendons.

Biological control systems have long been studied as possible inspiration for the construction of robotic controllers. In this paper, we present a control of voluntary movements using a cortical network within constraints from neurophysiology. Neural controller is proposed to control desired joint trajectories for multi-joint opponent muscle control of a robot finger. Each joint is controlled by an agonist-antagonist muscle pair. Neural model proposes functional roles for pre-central cortical cell types in the computation of a descending command to spinal alpha and gamma motoneurons. Neurons in anterior area 5 are proposed to compute the position of the link in question using corollary discharges and feedback from muscles spindles. Neurons in posterior area 5 use this position perception to compute a desired movement direction. Through experimental results, we showed that neural controller exhibits key kinematic properties of human movements, dynamics compensation and including asymmetric bell-shaped velocity profiles. Neural controller suggests how the brain may set automatic and volitional gating mechanisms to vary the balance of static and dynamic feedback information to guide the movement command and to compensate for external forces.

Automatical operation of space robots is a challenging and ultimate goal of space servicing. In this paper, we present a novel approach for tracking and catching operation of space robots based on learning and transferring human control strategies (HCS). We firstly use an efficient support vector machine (SVM) to parameterize the model of HCS, and then develop a new SVM-based leaning structure to improve HCS in tracking and capturing control. The approach is fundamentally valuable in dealing with some problems such as small sample data and local minima, which makes it efficient in modeling, understanding and transferring its learning process. The simulation results demonstrate that the proposed method is useful and feasible in tracking trajectory and catching objects autonomously.

Nowadays, mobile robots have attracted more and more attention from researchers due to their extensive applications. Mobile robots need to have the capabilities of autonomy and intelligence, and they pose a challenge to researchers, which is to design algorithms that allow the robots to function autonomously in unstructured, dynamic, partially observable, and uncertain environments [1]. Navigation is the key to the relative technologies of mobile robots and neural networks are widely used in the field of mobile robot navigation due to their properties such as nonlinear mapping, ability to learn from examples, good generalization performance, massively parallel processing, and capability to approximate an arbitrary function given sufficient number of neurons. This paper surveys the developments in the last few years of the neural networks with applications to mobile robot navigation.

Fault detection and diagnosis (FDD) are increasingly important for wheeled mobile robots (WMRs), especially those in unknown environments such as planetary exploration. There are many kinds of fault diagnosis methods available for mobile robots, including multiple model-based approaches, particle filter based approaches, sensor fusion based approaches. Currently, all of these methods are designed for complete models. However, completely modeling a system is difficult, even impossible. In this paper, particle filter and neural network are integrated to diagnose complex systems with imperfect models. Two features are extracted from particles: the sum of sample weights, and the maximal a posteriori probability. These features are further feed to a neural network to decide whether the estimation given by the particle filter is credible or not. An incredible estimation indicates that the true state isn’t included in the state space, i.e. it is a novel state (or an unknown fault). This method preserves the merits of particle filter and can diagnose known faults as well as detect unknown faults. It is testified on a real mobile robot.

Unmanned ground vehicles (UGVs) play an increasingly important role in future space exploration and battlefield. This work is concerned with the automatic path tracking control of UGVs. By using the structure properties of the system, neuro-adaptive control algorithms are developed for high precision tracking without involving complex design procedures – the proposed control scheme only demands partial information of the system, no detail description of the system model is needed. Furthermore, uncertain effects such as external disturbance and uncertain parameters can easily be handled. In addition, all the internal signals are uniformly bounded and the control torque is smooth anywhere.

In this study, ANNs are introduced to act as a bridge between detailed computer codes and compact simulators with an aim to improve the capabilities of compact expert system. The ANNs compensate for the inaccuracies of a compact expert system occurring from simplified governing equations and a reduced number of physical control volumes, and predict the critical parameter usually calculated from the sophisticated computer code. To verify the appli-cability of the proposed methodology, computer simulations are undertaken for loss of flow accidents (LOFA).

Short term load forecasting (STLF) has an essential role in the operation of electric power systems. Although artificial neural networks (ANN) based predictors are more widely used for STLF in recent years, there still exist some difficulties in choosing the proper input variables and selecting an appropriate architecture of the networks. A novel approach is proposed for STLF by combining mutual information (MI) and ANN. The MI theory is first briefly introduced and employed to perform input selection and determine the initial weights of ANN. Then ANN module is trained using historical daily load and weather data selected to perform the final forecast. To demonstrate the effectiveness of the approach, short-term load forecasting was performed on the Hang Zhou Electric Power Company in China, and the testing results show that the proposed model is feasible and promising for load forecasting.

In this paper a novel variant activation (transform) sigmoid function with three parameters is proposed, and then the improved BP algorithm based on it is educed and discussed, then Embedded Chaos-BP algorithm is proposed by means of combining the new fast BP algorithm and chaos optimization algorithm, Embedded chaos-BP algorithm converges fast and globally, and has no local minimum. The efficiency and advantage of our method is proved by simulation results of nonlinear function and prediction results of short-term load based on the improved and traditional BP ANNs.

Accurate Short term load forecasting (STLF) plays a significant role in the management of power system of countries and regions on the grounds of insufficient electric energy for increased need. This paper presents an approach of back propagation neural network with rough set (RSBP) for complicated STLF with dynamic and non-linear factors in order to develop the accuracy of predictions. Through attribution reduction based on variable precision with rough set, the influence of noise data and weak interdependency data to BP is avoided so the time taken for training is decreased. Using load time series from a practical system, we tested the accuracy of forecasting in specific days with comparison.

To overcome the defects of neural network (NN) using back-propagation algorithm (BPNN) such as slow convergence rate and easy to fall into local minimum, the particle swarm optimization (PSO) algorithm was adopted to optimize BPNN model for short-term load forecasting (SLTF). Since those defects are partly caused by the random selection of network’s initial values, PSO was used to optimize initial weights and thresholds of BPNN model, thus a novel model for STLF was built, namely PSO-BPNN model. The simulation results of daily and weekly loads forecasting for actual power system show that the proposed forecasting model can effectively improve the accuracy of SLTF and this model is stable and adaptable for both workday and rest-day. Furthermore, its forecasting performance is far better than that of simple BPNN model and BPNN model using genetic algorithm to determine the initial values (GA-BPNN).

Electric Power Load Forecasting is important for the economic and secure operation of power system, and highly accurate forecasting result leads to substantial savings in operating cost and increased reliability of power supply. Conventional load forecasting techniques, including time series methods and stochastic methods, are widely used by electric power companies for forecasting load profiles. However, their accuracy is limited under some conditions. In this paper, neural networks have been successfully applied to load forecasting. Forecasting model with Neural Networks is set up based on the analysis of the characteristics of electric power load, and it works well even with rapidly changing weather conditions. This paper also proposes a novel method to improve the generalization ability of the Neural Networks, and this leads to further increasing accuracy of load forecasting.

This paper proposes an artificial neural network (ANN) model to predict

m

-daily-ahead electricity price using direct forecasting approach on European Energy Exchange (EEX) market. The most important characteristic of this model is the single output node for

m

-period-ahead forecasts. The potentials of ANNs are investigated by employing cross-validation schemes. Out-of-sample performance evaluated with three criteria across five forecasting horizons shows that the proposed ANNs are more robust multi-step-ahead forecasting method than autoregressive error models (AUTOREG). Moreover, ANN predictions are quite accurate even when the length of forecast horizon is relatively short or long.

Market clearing price (MCP) is one of the most important factors impacting on power system. Taking into account the features of deregulation and fluctuation, this paper uses artificial neural network to forecast next-day MCP, with period-decoupled data sequence and wavelet transform. For the purpose of better performance, an improved learning algorithm of artificial fish-swarm is proposed. By simulating fish-swarm actions, in random searching for foods, artificial fish-swarm based neural network (AFNN) achieves global optimum. Comparing with traditional next-day MCP forecasting methods, the suggested method could achieve better adaptability and greater predictive accuracy, which was proved by the experimental results.

This paper presents an evolutionary neural network (ENN) approach for solving the power system unit commitment problem. The proposed ENN approach combines a genetic algorithm (GA) with a back-propagation neural network (BPNN). The BPNN is first used as a dispatch tool to generate raw unit combinations for each hour temporarily ignoring time-dependent constraints. Then, the proposed decoding algorithm decodes the raw committed schedule of each unit into a feasible one. The GA is then used to find the finally optimal schedule. The most difficult time-dependent minimal uptime/downtime constraints are satisfied throughout the proposed encoding and decoding algorithm. Numerical results from a 10-unit example system indicate the attractive properties of the proposed ENN approach, which are a highly optimal solution and faster rate of computation.

Adopting the fundamentals of PWM voltage source rectifier and the PID-control technology based on error-back-propagation neural network, an electronic power load simulator is designed in this paper, which can simulate the exact Volt-Ampere characteristics of power load and supply high quality feedback electric power. In order to enable the PID controller to perform better, genetic algorithms are used to accumulate the priori knowledge of the neural network’s connection weights, and the system voltages and the controlled parameters are forecasted. The experimental results show that the electronic power load simulator runs well when the Volt-Ampere characteristics simulated are time-variable or voltage disturbances occur in power system.

The distribution system problems, such as planning, loss minimization, and energy restoration, usually involve the phase balancing or network reconfiguration procedures. The determination of an optimal phase balance is, in general, a combinatorial optimization problem. This paper proposes optimal reconfiguration of the phase balancing using the neural network, to switch on and off the different switches, allowing the three phases supply by the transformer to the end-users to be balanced. This paper presents the application examples of the proposed method using the real and simulated test data.

A new multi-objective formulation named normalized weighting method combined with particle swarm optimization for the connections between distribution transformers and a primary feeder problem is presented. The performance of Particle Swarm Optimization can be improved by strategically selecting the starting positions of the particles by back-propagation neural network. Six important objectives are considered in this problem. These six objectives are of equal important to electric utility companies, but they are somewhat non-commensurable with each other. In view of this, a normalized weighting method for the multi-objective problem is proposed. It can provide a set of flexible solutions using particle swarm optimization by following the intention of decision makers. To increase the realism, the load and operating constraints of the system are all considered. Comparative studies on actual Tai-power systems are given to demonstrate the effectiveness of the phase load balancing and the improvement of operation efficiency for the proposed method.

Partial discharge (PD) pattern recognition is an important tool in HV insulation diagnosis. A PD pattern recognition approach of HV power transformers based on a neural network is proposed in this paper. A commercial PD detector is firstly used to measure the 3-D PD patterns of epoxy resin power transformers. Then, two fractal features (fractal dimension and lacunarity) extracted from the raw 3-D PD patterns are presented for the neural- network-based (NN-based) recognition system. The system can quickly and stably learn to categorize input patterns and permit adaptive processes to access significant new information. To demonstrate the effectiveness of the proposed method, the recognition ability is investigated on 150 sets of field tested PD patterns of epoxy resin power transformers. Different types of PD within power transformers are identified with rather encouraged results.

With the growing application of Static Var Compensator (SVC) and Static Synchronous Compensator (STATCOM), the coordinating problem of SVC and STATCOM controllers in joint operation must be considered in modern power systems. This paper firstly establishes the nonlinear differential-algebra equations model of a single-machine infinite-bus (SMIB) power system installed with a SVC and a STATCOM and points out the possibility of the negative interactions between SVC and STATCOM controllers in this SMIB power system. Hence, a self-adaptive single neuron (SSN) control approach based on genetic algorithm is designed to eliminate the negative interactions and improve the stability of the closed-loop SMIB power system. The detailed simulation results demonstrate the effectiveness of the proposed SSN controllers.

This paper presents a power system stabilizer design that uses a generalized minimum variance-inverse dynamic neuro controller, which is the combination of the inverse dynamic neural model, the generalized minimum variance, and the neuro compensator. An inverse dynamic neural model represents the inverse dynamics of the system. The inverse dynamic neural model is trained to provide control input into the system, which makes the plant output reach the target value at the next sampling time. Once the inverse dynamic neural model is trained, it does not require retuning for cases with other types of disturbances. In this paper, a generalized minimum variance control scheme is adapted to prevent unstable system performance caused by non-minimum phase characteristics. In addition, a neural compensator is designed to compensate for modeling errors. The proposed control scheme is tested in a multimachine power system.

Artificial neural networks can be used as intelligent controllers to control non-linear, dynamic systems through learning, which can easily accommodate the non-linearities and time dependencies. However, they require large training time and large number of neurons to deal with complex problems. Taking benefit of the characteristics of a Generalized Neuron that requires much smaller training data and shorter training time, the pseudo-linear neural network (PNN) based model predictive approach used in the single and multi-machine power system studies is proposed in this paper. A simulation is carried out. It is demonstrated that the proposed control strategy is applicable to some of nonlinear systems.

For enhancing approximation ability of chaotic neural network to nonlinear system, some characteristics are researched about neuron algorithm, architecture of network and learning rule of neural network. A local recurrent chaotic neural network is constructed based on Aihara chaotic neuron. A heuristic modified improved BP algorithm is applied in the chaotic neural network training with well ability of convergence and stability. The chaotic neural network is applied in marine generator modeling for a real time simulator. The application indicates that the chaotic neural network can be applied to build marine generator with ideal ergodicity and few number of neuron. There are relationships between value of mean square error and chaotic characteristic of neuron in marine generator modeling. When the neuron is in chaotic state, the minimum value of mean square error will be acquired.

Ship synchronous generator modeling is the basis of control, analysis and design in the ship power systems. According to the strong non-linear relation characteristics of ship synchronous generator, a dynamic modeling method based on rough set theory (RST) and radial basis function (RBF) neural networks is presented in this paper. With the advantage of finding useful and minimal hidden patterns in data, RST is first applied to intelligent data analysis in this algorithm, including incompatible data elimination, important input nodes selection and radial basis function centers initialization, followed by a second stage adjusting the network parameters and training the weights of hidden nodes. The experimental results proved that this method could achieve greater accuracy and generalization ability.

The paper analyzes the structure character of a wind power generation and flywheel energy storage combined system, and presents a new control strategy—fuzzy neural network (FNN) based on genetic arithmetic (GA) for the nonlinear problem of the control system. The control strategy realizes automatic regulation of direct current (DC) bus voltage in the system, so it stabilizes the DC bus voltage of the system. The experimental results show that the controller has better self-learning and robustness, and realizes the satisfactory operation of the proposed system.

Recognition of power quality events by analyzing the voltage and current waveform disturbances is a very important task for the power system monitoring. This paper presents a new approach for the recognition of power quality disturbances using wavelet transform and neural networks. The proposed method employs the wavelet transform using multiresolution signal decomposition techniques working together with multiple neural networks using a learning vector quantization network as a powerful classifier. Various transient events are tested, such as voltage sag, swell, interruption, notching, impulsive transient, and harmonic distortion show that the classifier can detect and classify different power quality signal types efficiency.

This paper presents a novel self-adaptive wavelet neural network method for automatic recognition and classification of power quality disturbances. The types of disturbances include harmonic distortions, flickers, voltage sags, voltage swells, voltage interruptions, voltage notches, voltage impulses and voltage transients. The self-adaptive wavelet neural network model constructed consists of four layers: input layer, preprocessing layer, hidden layer and output layer. The preprocessing layer is also called wavelet layer whose function is to extract features of power quality disturbances for recognition and classification; the other three layers just constitute the feedforward neural network whose function is to recognize and classify the types of power quality disturbances. The self-adaptive wavelet neural network has a good anti-interference performance, and the test and evaluation results demonstrate that utilizing it power quality disturbances can be recognized and classified effectively, accurately and reliably.

This paper proposes a novel neural network called Extension Neural Network-Type 1 (ENN1) for vibration fault recognition according to generator vibration characteristic spectra. The proposed ENN1 has a very simple structure and permits fast adaptive processes for new training data. Moreover, the learning speed of the proposed ENN1 is shown to be faster than the previous approaches. The proposed method has been tested on practical diagnostic records in China with rather encouraging results.

Through the analysis of function approximation with wavelet transformation, an adaptive wavelet neural network is introduced in the paper, which is applied in data compression of fault data in power system. In addition, the wavelet entropy is adopted to choose the hidden nodes in the wavelet neural network. The learning algorithm of the wavelet neural network based on wavelet entropy is proposed and discussed for data compression of fault data in power system. The simulation results show that it is feasible and valid in the end.

This paper proposes a hybrid neural network model based on the Generalized Learning Vector Quantization(GLVQ) learning algorithm and applies this proposed method to the BIT system of More-Electric Aircraft Electrical Power System (MEAEPS). This paper first discusses the feasibility of application unsupervised neural networks to the BIT system and the representative Generalized LVQ (GLVQ) neural network is selected due to its good performance in clustering analysis. Next, we adopt a new form of loss factor to modify the original GLVQ algorithm in order to make it more suitable for our application. Since unsupervised networks cannot distinguish the similar classes, we add a LVQ layer to the GLVQ network to construct a hybrid neural network model. Finally, the proposed method has been applied to the intelligent BIT system of the MEAEPS, and the results show that the proposed method is promising to improve the performance of the BIT system.

Static voltage security is one of the important items of power system security. This paper provides an approach to calculate risk of low voltage in power system using neural network ensemble. Risk is defined as a condition under which there is a possibility of an adverse deviation from a desired outcome that is expected or hoped for. Risk index is used as an indicator of the low voltage security. It is calculated as the product of the probability of contingency and the impact of low voltage. Neural network ensemble (NNE) is used for the low voltage risk assessment to get the desired speed, accuracy and efficiency. The New England 39-bus test system is used as an example to demonstrate the efficiency of the proposed algorithm.

This paper provides a strategy at the system level to assess and mitigate power system cascading outages considering the probabilities of hidden failure in protection system, which affect the risk of cascading outages. Some risk indices are used to assess the risk of cascading outages. The fuzzy neural network (FNN) is used to obtain the risk indices and to propose a solution that can decrease the system cascading outage risk under limited budget. The IEEE 118-bus system is used to illustrate the methodology.