Advances in Intelligent Computing

This paper suggests a new diffusion method, which based on modified coherence diffusion for the enhancement of ocular fundus images (OFI) and parallel AOS scheme is applied to speed algorithm, which is faster than usual approach and shows good performance. A structure tensor integrating the second-order directional differential information is applied to analyze weak edges, narrow peak, and vessels structures of OFI in diffusion. The structure tensor and the classical one as complementary descriptor are used to build the diffusion tensor. The several experiment results are provided and suggest that it is a robust method to prepare image for intelligent diagnosis and instruction for treatment of ocular diseases. The modified diffusion for the enhancement of OFI can preserve important oriented patterns, including strong edges and weak structures.

An iterative hybrid interpolation method is proposed in this study, which is an integration of the bilinear and the bi-cubic interpolation methods and implemented by an iterative scheme. First, the implement procedure of the iterative hybrid interpolation method is described. This covers (a) a low resolution image is interpolated by using the bilinear and the bi-cubic interpolators respectively; (b) a hybrid interpolated result is computed according to the weighted sum of both bilinear interpolation result and bi-cubic interpolation result and (c) the final interpolation result is obtained by repeating the similar steps for the successive two hybrid interpolation results by a recursive manner. Second, a further discussion on the method – the relation between hybrid parameter and details of an image is provided from the theoretical point of view, at the same time, an approach used for the determining of the parameter is proposed based on the analysis of error parameter curve. Third, the effectiveness of the proposed method is verified based on the experimental study.

The failure probability of a system can be expressed as an integral of the joint probability density function within the failure domain defined by the limit state functions of the system. Generally, it is very difficult to solve this integral directly. The evaluation of system reliability has been the active research area during the recent decades. Some methods were developed to solve system reliability analysis, such as Monte Carlo method, importance sampling method, bounding techniques and Probability Network Evaluation Technique (PNET). This paper presents the implementation of several optimization algorithms, modified Method of Feasible Direction (MFD), Sequential Linear Programming (SLP) and Sequential Quadratic programming (SQP), in order to demonstrate the convergence abilities and robust nature of the optimization technique when applied to series system reliability analysis. Examples taken from the published references were calculated and the results were compared with the answers of various other methods and the exact solution. Results indicate the optimization technique has a wide range of application with good convergence ability and robustness, and handle problems under generalized conditions or cases.

This paper proposes a novel registration algorithm based on Pseudo-Polar Fast Fourier Transform (FFT) and Analytical Fourier-Mellin Transform (AFMT) for the alignment of images differing in translation, rotation angle, and uniform scale factor. The proposed algorithm employs the AFMT of the Fourier magnitude to determine all the geometric transformation parameters with its property of the invariance to translation and rotation. Besides, the proposed algorithm adopt a fast high accuracy conversion from Cartesian to polar coordinates based on the pseudo-polar FFT and the conversion from the pseudo-polar to the polar grid, which involves only 1D interpolations, and obtain a more significant improvement in accuracy than the conventional method using cross-correlation. Experiments show that the algorithm is accurate and robust regardless of white noise.

In this paper, an algorithm is proposed for detecting texts in images and video frames. Firstly, it uses the variances and covariancs on the wavelet coefficients of different color channels as color textural features to characterize text and non-text areas. Secondly, the k-means algorithm is chosen to classify the image into text candidates and background. Finally, the detected text candidates undergo the empirical rules analysis to identify text areas and project profile analysis to refine their localization. Experimental results demonstrate that the proposed approach could efficiently be used as an automatic text detection system, which is robust for font-size, font-color, background complexity and language.

This paper presents a novel approach that uses the dynamic time warping (DTW) to match the crucial points of signatures. Firstly, the signatures are aligned through the DTW and the crucial points of signatures are matched according to the mapping between the signatures. Then the signatures are segmented at these matched crucial points and the comparisons are accomplished between these segments. Experimental results show that such a strategy is quite promising.

Based on the dual tree complex wavelet transform and edge detection, a SAR image despeckling algorithm is proposed. It can be used to remove white Gauss additive noise (WGAN) too. The DT-CWT has the properties of shift invariance and more directions. Edges are effectively extracted based on this complex transform and adjacent scales coefficients multiplication. According to the statistical property of the edge and non edge wavelet coefficients, Laplacian and Gaussian distribution are used to describe them respectively. Bayesian MAP estimator is used to estimate the noiseless wavelet coefficient values. Analysis and experiments illustrate the effectiveness of the proposed algorithm.

Mutual Information has been used as a similarity metric in medical images registration. But local extrema impede the registration optimization process and rule out the registration accuracy, especially for rotation registration. In this paper, a novel approach to rotate registration based on image symmetry measure is presented. Image symmetry measure is defined to measure the symmetry about the possible axis. The symmetry measure is at its maximum when the possible symmetry axis is the real symmetry axis. The angle between the symmetry axes of two images can be used to estimate rotate registration parameter in advance without translation parameter. This method is of great benefit to rotation registration accuracy and avoids the disadvantage of traditional MI method searching in the multi-dimensional parameter space. Experiments show that our method is feasible and effective to rotation registration of medical images, which have obvious symmetry characteristics.

We propose a new method to solve the occlusion problem efficiently in rigid target tracking by combining integral-intensity-matching algorithm with multi-block-voting algorithm. If the target is occluded, means some blocks are occluded and tracked falsely. Then we don’t let the occluded blocks participate in voting and integral-intensity-matching calculation, and use the remainder unoccluded blocks which can represent target’ attribute to track the target unceasingly. Experimental results show that the adopted two algorithms are complementary, and effective combination can achieve reliable tracking performance under heavy occlusion.

Recognizing plant leaves has so far been an important and difficult task. This paper introduces a method of recognizing leaf images based on shape features using a hypersphere classifier. Firstly, we apply image segmentation to the leaf images. Then we extract eight geometric features including rectangularity, circularity, eccentricity, etc, and seven moment invariants for classification. Finally we propose using a moving center hypersphere classifier to address these shape features. As a result there are more than 20 classes of plant leaves successfully classified. The average correct recognition rate is up to 92.2 percent.

As important premises of automatic seal verification system, candidate seal must be detected from processed image and done registration with the template seal. The paper gives such a robust method. After the candidate seal is detected from the processed image by contour skeleton analysis, FFT is performed for template and candidate seals. FFT Magnitude feature matrix describing global and invariant properties of the seal image is constructed by integrating the Fourier transformation over each of the regions of a wedge-ring-detector. Robust rotation angle is evaluated by minimizing the difference between two feature matrixes for the two seals. Then, relative translation can be evaluated by limited position enumerating in the space domain. Experiment results show that our method can deal with noise-corrupted images and complicate-background images. Seal detection and registration are fast and accurate, and the methods have been used in a real seal identification system successfully

.

Mean Squared Error (MSE) has been

the

performance metric in most performance appraisals up to date if not all. However, MSE is useful only if an original non degraded image is available in image restoration scenario. In blind image restoration, where no original image exists, MSE criterion can not be used. In this article we introduce a new concept of incorporating Human Visual System (HVS) into blind restoration of degraded images. Since the image quality is subjective in nature, human observers can differently interpret the same iterative restoration results. This research also attempts to address this problem by quantifying some of the evaluation criteria with significant improvement in the consistency of the judgment of the final result. We have modified some image fidelity metrics such as MSE, Correlation Value and Laplacian Correlation Value metrics to be used in iterative blind restoration of blurred images. A detailed discussion and some experimental results pertaining to these issues are presented in this article.

Dimension reduction is a crucial step for pattern recognition and information retrieval tasks to overcome the curse of dimensionality. In this paper a novel unsupervised linear dimension reduction method,

Neighborhood Preserving Projections

(NPP), is proposed. In contrast to traditional linear dimension reduction method, such as principal component analysis (PCA), the proposed method has good neighborhood-preserving property. The main idea of NPP is to approximate the classical locally linear embedding (i.e. LLE) by introducing a linear transform matrix. The transform matrix is obtained by optimizing a certain objective function. Preliminary experimental results on known manifold data show the effectiveness of the proposed method.

This paper presents a structured light pattern for moving objects sensing in dynamic scenes. The proposed binary pattern can be projected by laser illumination, which aims at eliminate the affect of ambient sunlight, so as to widen application fields of depth sending approach based on structure light. Without the help of color information, the binary can provide great number of code words to make all sub-pattern own a globe unique code to make it suitable for moving objects sensing at one-shot. The propose patter offers more measurement spots than traditional patterns based on M-array, so as to acquire higher resolution. In this paper, the proposed pattern and codification are firstly presented. A new algorithm for fractured contour searching and searching strategies are discussed. An algorithm based on angle variation for contour character identification is given. Code points mapping and mapping regulations are also presented.

A novel Linear Hashtable Method Predicted Hexagonal Search (LHMPHS) method for block based motion compensation is proposed. Fast block matching algorithms use the origin as the initial search center, which often does not track motion very well. To improve the accuracy of the fast BMA’s, we employ a predicted starting search point, which reflects the motion trend of the current block. The predicted search centre is found closer to the global minimum. Thus the center-biased BMA’s can be used to find the motion vector more efficiently. The performance of the algorithm is evaluated by using standard video sequences, considers the three important metrics: The results show that the proposed algorithm enhances the accuracy of current hexagonal algorithms and is better than Full Search, Logarithmic Search etc.

Based on the assumption that background figures have been extracted form the input image, we propose a method that can effectively detection the moving objects from image sequence in this paper. The background difference, background difference based neighborhood pixels and frame difference information are fused to get the seed points of real moving object, only the blobs in moving detection based on background difference that intersect with seed pixels are selected as the final moving segmentation result, then we can obtain the true moving foreground. Simulation results show that the algorithm can avoid the false detection due to the wrong in background model or background update and can handle situation where the background of the scene contains small motions, and motion detection and segmentation can be performed correctly.

In this paper, we proposed a novel entropy coding called bit-precision method. Huffman coding and arithmetic coding are among the most popular methods for entropy-coding the symbols after quantization in image coding. Arithmetic coding outperforms Huffman coding in compression efficiency, while Huffman coding is less complex than arithmetic coding. Usually, one has to sacrifice either compression efficiency or computational complexity by choosing Huffman coding or arithmetic coding. We proposed a new entropy coding method that simply defines the bit precision of given symbols, which leads to a comparable compression efficiency to arithmetic coding and to the lower computation complexity than Huffman coding. The proposed method was tested for lossless image coding and simulation results verified that the proposed method produces the better compression efficiency than (single model) arithmetic coding and the substantially lower computational complexity than Huffman coding.

In this paper, a new method based on wavelet package transform is proposed for classification of texture images. It has been demonstrated that a large amount of texture information of texture images is located in middle-high frequency parts of image, a corresponding method called wavelet package transform, not only decomposing image from the low frequency parts, but also from the middle-high frequency parts, is presented to segment texture images into a few texture domains used for image classification. Some experimental results are obtained to indicate that our method for image classification is superior to the co-occurrence matrix technique obviously.

In this paper, palmprint features were classified into Local Features and Global Features. Based on this definition, we discussed the advantage and weakness of each kind of features and presented a new palmprint identification algorithm using combination features. In this algorithm, a new method for capturing the key points of hand geometry was proposed. Then we described our new method of palmprint feature extracting. This method considered both the global feature and local detail of a palmprint texture and proposed a new kind of palmprint feature. The experimental results demonstrated the effectiveness and accuracy of these proposed methods.

The waveform interpolation is one of the speech coding algorithms with high quality at low bit rates. In the WI coding, the vector quantization of SEW requires a variable dimension quantization technique since the dimension of the SEW amplitude spectrum varies depending on the pitch period. However, since the variable dimension vector makes a difficulty to employ conventional vector quantization techniques directly, some dimension conversion techniques are usually utilized for the quantization of the variable dimension vectors. In this paper, we propose a new dimension conversion method for the SEW quantization in order to reduce the cost of codebook storage space with a small conversion error in the wideband WI speech coding. This dimension conversion method would be more useful for the wideband speech because wideband speech requires larger codebook memory for the variable dimension vector quantization compared to narrowband speech.

This paper proposes a brand-new preprocessing scheme using the ranking of co-occurrence count about indices in neighboring pixels. Original indices in an index image are substituted by their ranks. Arithmetic coding, then, is followed. Using this proposed algorithm, a better compression efficiency can be expected with higher data redundancy because the indices of the most pixels are concentrated to the relatively few rank numbers. Experimental results show that the proposed algorithm achieves a better compression performance up to 26–48% over GIF, arithmetic coding and Zeng’s scheme.

In front-view facial images, chin contour is a relative stable shape feature and can be widely used in face recognition. But it is hard to extract by conventional edge-detection methods due to the complexities of grayscale distribution in chin area. This paper presents an effective approach to chin contour extraction using the facial parts distributing rules and approved snake model. We first approximately localize a parabola as the initial contour according to prior statistical knowledge, then use approved active contour model to find the real chin contour through iteration. Experimental results show that by this algorithm we can extract the precise chin contour which preserves lots of details for face recognition.

In this paper, we propose a new face recognition method under arbitrary lighting conditions, given only a single registered image and training data under unknown illuminations. Our proposed method is based on the exemplars which are synthesized from photometric stereo images of training data and the linear combination of those exemplars are used to represent the new face. We make experiments for verifying our approach and compare it with two traditional approaches. As a result, higher recognition rates are reported in these experiments using the illumination subset of Max-Planck Institute Face Database.

We present a technique for low-dimensional representation of facial images that achieve graceful degradation of recognition performance. We have observed that if data is well-clustered into classes, features extracted from a topologically continuous transformation of the data are appropriate for recognition when low-dimensional features are to be used. Based on this idea, our technique is composed of two consecutive transformations of the input data. The first transformation is concerned with best separation of the input data into classes and the second focuses on the transformation that the distance relationship between data points before and after the transformation is kept as closely as possible. We employ FLD (Linear Discriminant Analysis) for the first transformation, and classical MDS (Multi-Dimensional Scaling) for the second transformation. We also present a nonlinear extension of the MDS by ‘kernel trick’. We have evaluated the recognition performance of our algorithms: FLD combined with MDS and FLD combined with kernel MDS. Experimental results using FERET facial image database show that the recognition performances degrade gracefully when low-dimensional features are used.

This paper proposes a new method of extracting and tracking a nonrigid object moving while allowing camera movement. For object extraction we first detect an object using watershed segmentation technique and then extract its contour points by approximating the boundary using the idea of feature point weighting. For object tracking we take the contour to estimate its motion in the next frame by the maximum likelihood method. The position of the object is estimated using a probabilistic Hausdorff measurement while the shape variation is modelled using a modified active contour model. The proposed method is highly tolerant to occlusion. Because the tracking result is stable unless an object is fully occluded during tracking, the proposed method can be applied to various applications.

An induction motor control system fed by an AC/DC rectifier and a DC/AC inverter group is a nonlinear and EMI generating load causes harmonic distortions and EMI noise effects in power control systems. In this paper, a simulation model is designed for the control circuit and the harmonic effects of the nonlinear load are investigated using FFT analyses. Also, the EMI noise generated by the switching-mode power electronic devices measured using high frequency spectrum scopes. An LISN based active filter is used to damp the harmonic distortions and EMI noises in the simulation environment. Neural network based control system is used to tune the power op-amp gain of series active filter to obtain the voltage value stability at the equipment side as well.

In this paper, a new approach for leaf recognition using the result of segmentation of leaf’s skeleton based on the combination of wavelet transform (WT) and Gaussian interpolation is proposed. And then the classifiers, a nearest neighbor classifier (1-NN), a

K

-nearest neighbor classifier (k-NN) and a radial basis probabilistic neural network (RBPNN) are used, based on run-length features (RF) extracted from the skeleton to recognize the leaves. Finally, the effectiveness and efficiency of the proposed method is demonstrated by several experiments. The results show that the skeleton can be successfully and obviously extracted from the whole leaf, and the recognition rates of leaves based on their skeleton can be greatly improved.

A novel method of designing efficient SVM for fast color image segmentation is proposed in this paper. For application of large-scale image data, a new approach to initializing training set via pre-selecting useful training samples is adopted. By using a morphological unsupervised clustering technique, samples at the boundary of each cluster are selected for SVM training. With the proposed method, various experiments are carried out on the color blood cell images. Results show that the training set and time can be decreased considerably without lose of any segmentation accuracy.

Existing methods to deal with off-line signature verification usually adopt the feature representation based approaches which suffer from limited training samples. It is desired to employ straightforward means to measure similarity between 2-D static signature graphs. In this paper, we incorporate merits of both global and local alignment methods. Two signature patterns are globally registered using weak affine transformation and correspondences of feature points between two signature patterns are determined by applying an elastic local alignment algorithm. Similarity is measured as the mean square of sum Euclidean distances of all found corresponding feature points based on a match list. Experimental results showed that the computed similarity measurement was able to provide sufficient discriminatory information. Verification performance in terms of equal error rate was 18.6% with four training samples.

In this paper an efficient shape matching and recognition approach based on genetic algorithm is proposed and successfully applied to plant special identification. Firstly, a Douglas-Peucker approximation algorithm is adopted to the original shape and a new shape representation is used to form the sequence of invariant attributes. Then a genetic algorithm for shape matching is proposed to do the shape recognition. Finally, the superiority of our proposed method over traditional approaches to plant species identification is demonstrated by experiment. The experimental result showed that our proposed genetic algorithm for leaf shape matching is much suitable for the recognition of not only intact but also blurred, partial, distorted and overlapped plant leaves due to its robustness.

In this paper, we provide a steganalysis method which can detect the hiding in the least significant bit of the DCT coefficients. The method is based on the thought that the DCT coefficients are correlative. So the LSB sequence of the DCT coefficients is not random as a pseudo-random sequence. The randomness the LSB sequence is measured by some statistical tests. We find, as the increase of the embedded secrets, the randomness of the LSB sequence increase. Using the statistical tests as the features, we train

ε

-support vector regression (

ε

-SVR) with train images to get the statistical mode of the estimation of the embed secrets. With the statistical mode, we can discriminate the stego-images from the clear ones. We test our method on Jsteg and OutGuess. The results of experiments show that our method can detect the hiding by Jsteg and OutGuess either.

People are difficult targets to process in video surveillance and monitoring (VSAM) because of small size and non-rigid motion. In this paper, we address neural network application to people tracking for VSAM. A feedforward multilayer perceptron network (FMPN) is employed for the tracking in low-resolution image sequences using position, shape, and color cues. When multiple people are partly occluded by themselves, the foreground image patch of the people group detected is divided into individuals using another FMPN. This network incorporates three different techniques relying on a line connecting top pixels of the binary foreground image, the vertical projection of the binary foreground image, and pixel value variances of divided regions. The use of neural networks provides efficient tracking in real outdoor situations particularly where the detailed visual information of people is unavailable due mainly to low image resolution.

How face pose (rotating from right to left) influence the fusion authentication accuracy in face and fingerprint identity authentication system? This paper firstly tries to answer this question. The maximum rotating degree that fusion system can bear is given out by experiment. Furthermore, theoretical analysis deals with how face pose influence the fusion performance is proposed in this paper. Experiment results show that faces with big rotated degree can not be helpful but harmful to fusion system. And the maximum rotated angle of face that fusion system can bear is 20 degree. On the other hand, theoretical analysis proved that the mathematical inherence of influence of face pose on fusion system is not only the reduction of variance but also the decrease of distance between the genuine and imposter classes.

In this paper, a novel method of blind Super-Resolution (SR) image restoration is presented. First, a learning based blur identification method is proposed to identify the blur parameter in which Sobel operator and Vector Quantization (VQ) are used for extracting feature vectors. Then a super-resolution image is reconstructed by a new hybrid MAP/POCS method where the data fidelity term is minimized by

l

1

norm and regularization term is defined on the high frequency sub-bands offered by Stationary Wavelet Transform (SWT) to incorporate the smoothness of the discontinuity field. Simulation results demonstrate the effectiveness and robustness of our method.

In this paper, we model occlusion and appearance/disappearance in multi-target tracking in video by three coupled Markov random fields that model the following: a field for joint states of multi-target, one binary process for existence of individual target, and another binary process for occlusion of dual adjacent targets. By introducing two robust functions, we eliminate the two binary processes, and then apply a novel version of belief propagation called sequential stratified sampling belief propagation algorithm to obtain the maximum a posteriori (MAP) estimation in the resulted dynamic Markov network. By using stratified sampler, we incorporate bottom-up information provided by a learned detector (e.g. SVM classifier) and belief information for the messages updating. Other low-level visual cues (e.g. color and shape) can be easily incorporated in our multi-target tracking model to obtain better tracking results. Experimental results suggest that our methods are comparable to the state-of-the-art multiple targets tracking methods in several test cases.

The techniques for organizing and retrieving the artifacts from digital libraries (DLs) semantically are discussed, which include letting the taxonomies and semantic relations work in tandem to index the artifacts in DLs; integrating the techniques used in natural language processing and taxonomies to help users to start their retrieval processes; and ranking scientific papers on similarity in terms of contents or ranking the relevant papers on multi-factors. These techniques are verified through the design and implementation of a prototype of DLs for scientific paper management.

This paper is concerned with solution of the consistent fundamental matrix estimation in a quadratic measurement error model. First an extended system for determining the estimator is proposed, and an efficient implementation for solving the system-a continuation method is developed to fix on an interval in which a local minimum belongs. Then an optimization method using a quadratic interpolation is used to exactly locate the minimum. The proposed method avoids solving total eigenvalue problems. Thus the computational cost is significantly reduced. Synthetic and real images are used to verify and illustrate the effectiveness of the proposed approach.

Measurement of traffic time series plays a key role in the research of communication networks though theoretic research has a considerable advances. Differing from analytical analysis, quantities of interest are estimates experimentally analyzed from measured real life data. Hence, accuracy should be taken into account from a view of engineering. In practical terms, it is inappropriate to record data series that is either too short or over-long as too short record may not provide enough data to achieve a given degree of accuracy of an estimate while over-long record is usually improper for real-time applications. Consequently, error analysis based on record length has practical significance. This paper substantially extends our previous work [20,21] by detailing the derivations of error bound relating to record length and the Hurst parameter of a long-range dependent fractional Gaussian noise and by interpreting the effects of long-range dependence on record length. In addition, a theoretical evaluation of some widely used traces in the traffic research is also given.

Finding association rules is an important data mining problem and can be derived based on mining large frequent candidate sets. In this paper, a new algorithm for efficient generating large frequent candidate sets is proposed, which is called Matrix Algorithm. The algorithm generates a matrix which entries 1 or 0 by passing over the cruel database only once, and then the frequent candidate sets are obtained from the resulting matrix. Finally association rules are mined from the frequent candidate sets. Numerical experiments and comparison with the Apriori Algorithm are made on 4 randomly generated test problems with small, middle and large sizes. Experiments results confirm that the proposed algorithm is more effective than Apriori Algorithm.

Partner selection is a very popular problem in the research of virtual organization and supply chain management, the key step in the formation of virtual enterprise is the decision making on partner selection. In this paper, a activity network based multi-objective partner selection model is put forward. Then a new heuristic algorithm based on particle swarm optimization(PSO) and simulated annealing(SA) is proposed to solve the multi-objective problem. PSO employs a collaborative population-based search, which is inspired by the social behavior of bird flocking. It combines local search(by self experience) and global search(by neighboring experience), possessing high search efficiency. SA employs certain probability to avoid becoming trapped in a local optimum and the search process can be controlled by the cooling schedule. The hybrid algorithm combines the high speed of PSO with the powerful ability to avoid being trapped in local minimum of SA. We compare the hybrid algorithm to both the standard PSO and SA models, the simulation results show that the proposed model and algorithm are effective.

This paper proposed a modified algorithm, sequential niching particle swarm optimization (SNPSO), for the attempt to get multiple maxima of multimodal function. Based on the sequential niching technique, our proposed SNPSO algorithm can divide a whole swarm into several sub-swarms, which can detect possible optimal solutions in multimodal problems sequentially. Moreover, for the purpose of determining sub-swarm’s launch criteria, we adopted a new PSO space convergence rate (SCR), in which each sub-swarm can search possible local optimal solution recurrently until the iteration criteria is reached. Meanwhile, in order to encourage every sub-swarm flying to a new place in search space, the algorithm modified the raw fitness function of the new launched sub-swarm. Finally, the experimental results show that the SNPSO algorithm is more effective and efficient than the SNGA algorithm.

In this paper, we formally define the problem of outlier detection in categorical data as an optimization problem from a global viewpoint. Moreover, we present a local-search heuristic based algorithm for efficiently finding feasible solutions. Experimental results on real datasets and large synthetic datasets demonstrate the superiority of our model and algorithm.

In this paper, we implement a pilot study on the detection of abnormal financial asset trading activities using an artificial immune system. We develop a prototype

artificial immune abnormal-trading-detecting system

(AIAS)to scan the proxy data from the stock market and detect the abnormal trading such as insider trading and market manipulation, etc. among them. The rapid and real time detection capability of abnormal trading activities has been tested under simulated stock market as well as using real intraday price data of selected Australian stocks. Finally, three parameters used in the AIAS are tested so that the performance and robustness of the system are enhanced.

In this paper, we formulate the philosophy of Quantum-behaved Particle Swarm Optimization (QPSO) Algorithm, and suggest a parameter control method based on the population level. After that, we introduce a diversity-guided model into the QPSO to make the PSO system an open evolutionary particle swarm and therefore propose the Adaptive Quantum-behaved Particle Swarm Optimization Algorithm (AQPSO). Finally, the performance of AQPSO algorithm is compared with those of Standard PSO (SPSO) and original QPSO by testing the algorithms on several benchmark functions. The experiments results show that AQPSO algorithm outperforms due to its strong global search ability, particularly in the optimization problems with high dimension.

In this paper, we aim to develop a systematic framework to semi-automate the process of system logs and databases of intrusion detection systems (IDS). We use both Ef-attribute based mining and Es-attribute based mining to mine effective and essential attributes (hence interesting patterns) from the vast and miscellaneous system logs and IDS databases.

For the purpose of extending the Web that is able to understand and process information by machine, Semantic Web shared knowledge in the ontology form. For exquisite query processing, this paper proposes a method to use semantic relations in the ontology as relevance feedback information to query expansion. We made experiment on pharmacy domain. And in order to verify the effectiveness of the semantic relation in the ontology, we compared a keyword based document retrieval system that gives weights by using the frequency information compared with an ontology based document retrieval system that uses relevant information existed in the ontology to a relevant feedback. From the evaluation of the retrieval performance, we knew that search engine used the concepts and relations in ontology for improving precision effectively. Also it used them for the basis of the inference for improvement the retrieval performance.

An novel adaptive filtering method based on the wavelet transform is presented for a fiber optical gyroscope (FOG) on the moving base. Considering the performance difference of a FOG in different angular velocity, threshold values of different scales of wavelet coefficients are adjusted according to magnitude of FOG output signal, soft thresholding method is used to evaluate the wavelet coefficients, so effects of random signal noise and non-line of calibration factors of a FOG are removed at the maximum extent, and sensitivity of a FOG can be ensured. Filtering results of actual FOG show the proposed method has fine dynamic filtering effect.

Automatic text categorization is defined as the task to assign free text documents to one or more predefined categories based on their content. Classical method for computing text similarity is to calculate the cosine value of angle between vectors. In order to improve the categorization performance, this paper puts forward a new algorithm to compute the text similarity based on standard deviation. Experiments on Chinese text documents show the validity and the feasibility of the standard deviation-based algorithm.

ISOMap is a popular method for nonlinear dimensionality reduction in batch mode, but need to run its entirety inefficiently if the data comes sequentially. In this paper, we present an extension of ISOMap, namely I-ISOMap, augmenting the existing ISOMap framework to the situation where additional points become available after initial manifold is constructed. The MDS step, as a key component in ISOMap, is adapted by introducing Spring model and sampling strategy. As a result, it consumes only linear time to obtain a stable layout due to the Spring model’s iterative nature. The proposed method outperforms earlier work by Law [1], where their MDS step runs within quadratic time. Experimental results show that I-ISOMap is a precise and efficient technique for capturing evolving manifold.

The number of accessible Web pages has been growing fast on the Internet. It has become increasingly difficult for users to find information on the Internet that satisfies their individual needs. This paper proposes a novel approach and presents a prototype system for personalized information retrieval based on user profile. In our system, we return different searching results to the same query according to each user’s profile. Compared with other personalized search systems, we learn the user profile automatically without any effort from the user. We use the method of support vector machine to construct user profile. A profile ontology is introduced in order to standardize the user profile and the raw results returned by the search engine wrapper. Experiments show that the precision of the returned web pages is effectively improved.

In this paper, we propose a taxonomy for knowledge-oriented question, and study the machine learning based classification for knowledge-oriented Chinese questions. By knowledge-oriented questions, we mean questions carrying information or knowledge about something, which cannot be well described by previous taxonomies. We build the taxonomy after the study of previous work and analysis of 6776 Chinese knowledge-oriented questions collected from different realistic sources. Then we investigate the new task of knowledge-oriented Chinese questions classification based on this taxonomy. In our approach, the popular SVM learning method is employed as classification algorithm. We explore different features and their combinations and different kernel functions for the classification, and use different performance metrics for evaluation. The results demonstrate that the proposed approach is desirable and robust. Thorough error analysis is also conduced.

Tree augmented Naive Bayes (TAN) classifier is a good tradeoff between the model complexity and learnability in practice. Since there are few complete datasets in real world, in this paper, we develop research on how to efficiently learn TAN from incomplete data. We first present an efficient method that could estimate conditional Mutual Information directly from incomplete data. And then we extend basic TAN learning algorithm to incomplete data using our conditional Mutual Information estimation method. Finally, we carry out experiments to evaluate the extended TAN and compare it with basic TAN. The experimental results show that the accuracy of the extended TAN is much higher than that of basic TAN on most of the incomplete datasets. Despite more time consumption of the extended TAN compared with basic TAN, it is still acceptable. Our conditional Mutual Information estimation method can be easily combined with other techniques to improve TAN further.

Increasing the smooth degree of data series is key factor of grey model’s precision. In this paper, the more general method is put forward on the basis of summarizing several kinds of ways to improve smooth degree of data series, and a new transformation is represented. The practical application shows the effectiveness and superiority of this method.

In this paper, a general Combinatorial Ant System-based fault tolerant distributed routing algorithm modeled like a dynamic combinatorial optimization problem is presented. In the proposed algorithm, the solution space of the dynamic combinatorial optimization problem is mapped into the space where the ants will walk, and the transition probability and the pheromone update formula of the Ant System is defined according to the objective function of the communication problem.

The rapid growth of the World-Wide Web poses unprecedented scaling challenges for general-purpose crawlers. Topic-specific web crawler is developed to collect relevant web pages of interested topics form the Internet. Based on the analyses of HITS algorithm, a new P-HITS algorithm is proposed for topic-specific web crawler in this paper. Probability is introduced to select the URLs to get more global optimality, and the metadata of hyperlinks is appended in this algorithm to predict the relevance of web pages better. Experimental results indicate that our algorithm has better performance.

Elliptical flexure hinges are one of the most widely used flexure hinges for its high flexibility. To design elliptical flexure hinges of best performance, the author proposed a modified particle swarm optimization (MPSO) search method, where an exponentially decreasing inertia weight is deployed instead of a linearly decreasing inertia weight. Simulations indicate that the MPSO method is very effective. The optimal design parameters including the cutout configuration and the minimum thickness are obtained.

Kernel methods such as support vector machines have been used extensively for various classification tasks. In this paper, we describe an entropy based string kernel and a novel logistic kernel partial least square algorithm for classification of sequential data. Our experiments with a human chromosome dataset show that the new kernel can be computed efficiently and the algorithm leads to a high accuracy especially for the unbalanced training data.

This paper introduces an adaptive algorithm determining the measurement-track association problem in multi-target tracking. We model the target and measurement relationships and then define a MAP estimate for the optimal association. Based on this model, we introduce an energy function defined over the measurement space, that incorporates the natural constraints for target tracking. To find the minimizer of the energy function, we derived a new adaptive algorithm by introducing the Lagrange multipliers and local dual theory. Through the experiments, we show that this algorithm is stable and works well in general environments.

In this paper, we propose a novel approach to DNA computing- inspired semantic model. The model is theoretically proposed and constructed with DNA molecules. The preliminary experiment on construction of the small test model was successfully done by using very simple techniques: parallel overlap assembly (POA) method, polymerase chain reaction (PCR), and gel electrophoresis. This model, referred to as ‘

semantic model based on molecular computing’

(SMC) has the structure of a graph formed by the set of all (attribute, attribute values) pairs contained in the set of represented objects, plus a tag node for each object. Each path in the network, from an initial object-representing tag node to a terminal node represents the object named on the tag. Input of a set of input strands will result in the formation of object-representing dsDNAs via parallel self-assembly, from encoded ssDNAs representing both attributes and attribute values (nodes), as directed by ssDNA splinting strands representing relations (edges) in the network. The proposed model is very suitable for knowledge representation in order to store vast amount of information with high density.

Particle swarm optimization is a computational intelligence method of solving the multiobjective optimization problems. But for a given particle, there is no effective way to select its globally optimal particle and locally optimal particle. The particle angle is defined by the particle’s objective vector. The globally optimal particle is selected according to the minimal particle angle. Updating the locally optimal particle and particle swarm is based on the Pareto dominance relationship between the locally optimal particle and the offspring particles and the particle’s density. A multiobjective particle swarm optimization based on the minimal particle angle is proposed. The algorithm proposed is compared with sigma method ,NSPSO method and NSGA-II method on four complicated benchmark multiobjective function optimization problems. It is shown from the results that the Pareto front obtained with the algorithm proposed in this paper has good distribution, approach and extension properties.

An information fusion algorithm based on the quantum neural networks is presented for fault diagnosis in an integrated circuit. By measuring the temperature and voltages of circuit components of mate changing circuit board of photovoltaic radar, the fault membership functional assignment of two sensors to circuit components is calculated, and the fusion fault membership functional assignment is obtained by using the 5-level transfer function quantum neural network (QNN). Then the fault component is precisely found according to the fusion data. Comparing the diagnosis results based on separate original data DS fusion data BP fusion data with the ones based on QNN fused data, it is shown that the quantum fusion fault diagnosis method is more accurate.

Performance of speech emotion recognition largely depends on the acoustic features used in a classifier. This paper studies the statistical feature selection problem in Mandarin speech emotion recognition. This study was based on a speaker dependent emotional mandarin database. Pitch, energy, duration, formant related features and some velocity information were selected as base features. Some statistics of them consisted of original feature set and full stepwise discriminant analysis (SDA) was employed to select extracted features. The results of feature selection were evaluated through a LDA based classifier. Experiment results indicate that pitch, log energy, speed and 1st formant are the most important factors and the accuracy rate increases from 63.1 % to 76.5 % after feature selection. Meanwhile, the features selected by SDA are better than the results of other feature selection methods in a LDA based classifier and SVM. The best performance is achieved when the feature number is in the range of 9 to 12.

This paper presents a novel method for reconstructing 3D human body pose from monocular image sequences based on top-down learning. Human body pose is represented by a linear combination of prototypes of 2D silhouette images and their corresponding 3D body models in terms of the position of a predetermined set of joints. With a 2D silhouette image, we can estimate optimal coefficients for a linear combination of prototypes of the 2D silhouette images by solving least square minimization. The 3D body model of the input silhouette image is obtained by applying the estimated coefficients to the corresponding 3D body model of prototypes. In the learning stage, the proposed method is hierarchically constructed by classifying the training data into several clusters recursively. Also, in the reconstructing stage, the proposed method hierarchically reconstructs 3D human body pose with a silhouette image or a silhouette history image. We use a silhouette history image and a blurring silhouette image as the spatio-temporal features for reducing noise due to extraction of silhouette image and for extending the search area of current body pose to related body pose. The experimental results show that our method can be efficient and effective for reconstructing 3D human body pose.

This paper presents a database of 14 representative gestures in daily life of 20 subjects. We call this database the 2D and 3D Full-Body Gesture (FBG) database. Using 12 sets of 3D motion cameras and 3 sets of stereo cameras, we captured 3D motion data and 3 pairs of stereo-video data at 3 different directions for each gesture. In addition to these, the 2D silhouette data is synthesized by separating a subject and background in 2D stereo-video data and saved as binary mask images. In this paper, we describe the gesture capture system, the organization of database, the potential usages of the database and the way of obtaining the FBG database. We expect that this database would be very useful for the study of 2D/3D human gestures.

The capabilities and robustness of a new spiking neural network (SNN) learning algorithm are demonstrated with sound classification and function approximation applications. The proposed SNN learning algorithm and the radial basis function (RBF) learning method for function approximation are compared. The complexity of the learning algorithm is analyzed.

Writing forces are important dynamics of online signatures and they are harder to be imitated by forgers than signature shape. An improved DTW (Dynamic Time Warping) algorithm is put forward to verify online signatures based on writing forces. Compared to the general DTW algorithm, this one deals with the varying consistency of signature point, signing duration and the different weights of writing forces in different direction. The iterative dexperiment is introduced to decide weights for writing forces in different direction and the classification threshold. A signature database is constructed with F_Tablet and the equal error rate of 1.4% is realized with the improved algorithm.

Projective reconstruction is known to be an important step for 3D reconstruction in Euclidean space. In this paper, we present a new projective reconstruction algorithm based on invariant properties of the line segments in projective space: collinearity, order of contact, intersection. Points on each line segment in the image are reconstructed in projective space, and we determine the best-fit 3D line from them by Least-Median-Squares (LMedS). Our method regards the points unsatisfying collinearity as outliers, which are caused by false feature detection and tracking. In addition, both order of contact and intersection in projective space are considered. By using the points that are the orthogonal projection of outliers onto the 3D line, we iteratively obtain more precise projective matrix than the previous method.

In this paper, we present an ANN hybrid ensemble scheme for simultaneous object recognition and pose estimation from 2D multiple-view image sequence, and realized human vision simulation within an intelligent machine. Based on the notion of similarity measure at various metrics, the paradox between information simplicity and accuracy is balanced by a model view generation procedure. An ANN hierarchical hybrid ensemble framework, much like a decision tree, is then set up, with multiple weights and radial basis function neural networks respectively employed for different tasks. The strategy adopted not only determines object identity by spatial geometrical cognition and omnidirectional accumulation through connectivity, but also assigns an initial pose estimation on a viewing sphere in a coarse to fine process. Simulation experiment has achieved encouraging results, proved the approach effective, superior and feasible in large-scale database and parallel computation.

Intensity degradations are a familiar problem for fluorescein angiogram sequences. In this paper, we attempt to super-resolve a fluorescein angiogram, and to keep the high intensity pixels from degrading. To this end, we incorporate a new constraint, called intensity constraint, to Miller’s regularization formulation with a smoothness constraint. Considering the specified requirement for fluorescein angiograms, we also modify the

Q

-th order converging algorithm for implementation purpose. In our scheme, including its formulation and implementation, super-resolution reconstruction can not only handle the traditional problems, such as blur, decimation, and noise, but also achieve an important feature, intensity preservation. The experiments show that our approach has satisfactory results in the two aspects.

In this paper, we propose a novel on-line handwritten signature verification method. Firstly, the pen-position parameters of the on-line signature are decomposed into multiscale signals by wavelet transform technique. For each signal at different scales, we can get a corresponding zero-crossing representation. Then the distances between the input signature and the reference signature of the corresponding zero-crossing representations are computed as the features. Finally, we build a binary Support Vector Machine (SVM) classifier to demonstrate the advantages of the multiscale zero-crossing representation approach over the previous methods. Based on a common benchmark database, the experimental results show that the average False Rejection Rate (FRR) and False Acceptance Rate (FAR) are 5.25% and 5%, respectively, which illustrates such new approach to be quite effective and reliable.

Hand tracking is a challenging problem due to the complexity of searching in a 20+ degrees of freedom space for an optimal estimate. This paper develops a statistical method for robust visual hand tracking, in which graphical model decoupling different hand joints is performed to represent the hand constraints. Each node of the graphical model represents the position and the orientation of each hand joint in world coordinate. Then, the problem of hand tracking is transformed into an inference of graphical model. We extend Nonparametric Belief Propagation to a sequential process to track hand motion. The Experiment results show that this approach is robust for 3D hand motion tracking.

Identifying centerlines of vessels in the retinal image is helpful to provide useful information in diagnosis of eye diseases and early signs of systemic disease. This paper presents a novel thinning method based on the wavelet transform with a multilevel scheme. The development of the method is inspired by the favorable characteristics of wavelet transform moduli. Mathematical analysis is given to show that the vessel edge and centerline can be detected efficiently by computing the maxima and minima of wavelet transform moduli. The implementation is performed by applying various scale sizes of the wavelet transform to thin the multiple-pixel-wide ribbon-like vessel gradually to be one-pixel-wide centerline. Experiment results show that the identified centerline of vascular trees are accurate by visual inspection and are useful for further applications.

To describe the complicated neural dynamics of cerebra with time delays, a new type of model called generalized cellular neutral neural networks (GCNNNs) is studied in this paper. It is noted that the GCNNNs reduce to generalized cellular neural networks (GCNNs) in the absence of the neutral term in systems. Some criteria for mean square exponential stability and asymptotic stability of GCNNNs are established and the relationship between the neutral item and the whole system is analyzed. Simulation results are given to show the effectiveness of the proposed analysis algorithms.

One of the intelligent aspects of human beings in pattern recognition is that man identifies an object in real world using Marked Characteristic Principle (MCP). This paper proposes a humanoid recognition method for radar emitter signals. The main points of the method include feature ordering and an improved one-versus-rest multiclass classification support vector machines. According to MCP, an approach for computing marked characteristic coefficients is presented to obtain the most marked feature of every radar emitter signal. Subsequently, a support vector network is designed using the improved one-versus-rest combination approach of several binary support vector machines. Experimental results show that the introduced method has faster recognition speed and better classification capability than conventional recognition approaches.

This paper proposes two methods which take advantage of

k

-mean clustering algorithm to decrease the number of support vectors (SVs) for the training of support vector machine (SVM). The first method uses

k

-mean clustering to construct a dataset of much smaller size than the original one as the actual input dataset to train SVM. The second method aims at reducing the number of SVs by which the decision function of the SVM classifier is spanned through

k

-mean clustering. Finally, Experimental results show that this improved algorithm has better performance than the standard Sequential Minimal Optimization (SMO) algorithm.

This work analyses a recently proposed statistically based technique for monitoring complex dynamic process systems [17]. The technique utilises a state space model that is cast into the multivariate statistical process control framework (i) to define a set of state variables that can describe dynamic process behaviour, (ii) to generate univariate statistics that can monitor dynamic process behaviour and (iii) to construct contribution plots from these statistics that can diagnose anomalous process behaviour. The presented analysis reveals that the size of the state space monitoring model can be reduced. The utility of the improved dynamic monitoring technique is demonstrated using an industrial application study to a glass-melter process.

In many regression applications, there exist common cases for users to know qualitatively, yet partially, about nonlinear relationships of physical systems. This paper presents a novel direction for constructing feedforward neural networks (

FNNs

) which are subject to the given nonlinear relationships. The “

Integrated models

”, associating FNNs with the given nonlinear functions, are proposed. Significant benefits will be obtained over the conventional FNNs by using these models. First, they add a certain degree of comprehensive power for nonlinear approximators. Second, they may provide better generalization capabilities. Two issues are discussed about the improved approximation and the estimation of the real parameters to the partially known function in the proposed models. Numerical studies are given in comparing with the conventional FNNs.

This paper proposes a new image compression algorithm which combines SVM regression with wavelet transform. Compression is achieved by using SVM regression to approximate wavelet coefficients. Based on the characteristic of wavelet decomposition, the coefficient correlation in wavelet domain is analyzed. According to the correlation characteristic at different scales and orientations, three kinds of arranging methods of wavelet coefficients are designed, which make SVM compress the coefficients more efficiently. Moreover, an effective entropy coder based on run-length and arithmetic coding is used to encode the support vectors and weights. Experimental results show that the compression performance of the algorithm achieve much improvement.

In this paper, a new algorithm IPPSO (Parallel Particle Swarm Optimization with Island model) is proposed. It aims at remedying the defect of superquadric parametric fitting problem which is solved with L-M (Levenberg- Marquardt) method in 3D reconstruction and improving the algorithm performance of particle swarm optimization for application to large-scale problems and multi-variable solutions. This paper investigates 3D representation characteristics of superquadrics and makes analysis for the defect of superquadric parametric model fitting by L-M algorithm. It presents the principle and the implementation of superquadric parametric model fitting by using IPPSO. In addition, it describes the design principle and implementation method of IPPSO. In the end, the simulation results are analyzed. The results show the good effectiveness of the proposed approach, especially in the accuracy and discernment of superquadric 3D models reconstruction for objects.

A control strategy for precision position tracking of the magnetostrictive actuator (MA) with dominant hysteresis is proposed. In this strategy, a dynamic recurrent neural network with hysteron (DRNNH) is adopted as a feedforward controller for on-line learning the inverse model of the MA to remove the effect of the hysteresis of the MA. A proportional-plus-derivative (PD) feedback controller is used to reduce the position tracking error. Simulation results validate the excellent performances of the control strategy.

An orthogonal forward selection (OFS) algorithm based on the leave-one-out (LOO) criterion is proposed for the construction of radial basis function (RBF) networks with tunable nodes. This OFS-LOO algorithm is computationally efficient and is capable of identifying parsimonious RBF networks that generalise well. Moreover, the proposed algorithm is fully automatic and the user does not need to specify a termination criterion for the construction process.

This paper presents a novel recurrent time continuous neural network model for solving eigenvalue and eigenvector problem. The network is proved to be globally convergent to an exact eigenvector of a matrix

A

with respect to the problem’s feasible region. This convergence is called quasi-convergence in the sense of the starting point to be in the feasible set. It also demonstrates that the network is primal in the sense that the network’s neural trajectories will never escape from the feasible region when starting at it. By using an energy function, the network’s stable point set is guaranteed to be the eigenvector set of the involved matrix. Compared with the existing neural network models for eigenvalue problem, the new model’s performance is more effective and more reliable. Moreover, simulation results are given to illustrate further the global convergence and the fundamental validity of the proposed neural network for eigenvalue problem.

This paper investigates the chaos synchronization problem of a new 4-scroll chaotic system. Three nonlinear control approaches via state variables are studied, namely nonlinear feedback control, adaptive control and adaptive sliding mode type of control. Based on Lyapunov stability theory, control laws are derived such that the two identical 4-scroll systems are to be synchronized. Some sufficient conditions for the synchronization are obtained analytically in three cases. Numerical simulation results are given to show the effectiveness of the proposed methods.

In this paper, the requirement of Lipschitz condition on the activation functions is essentially dropped. By using Lyapunov functional and Young inequality, some new criteria concerning global exponential stability are obtained for generalized neural networks with variable coefficients and distributed delays. Since these new criteria do not require the activation functions to be differentiable, bounded or monotone nondecreasing and the connection weight matrices to be symmetric, they are mild and more general than previously known criteria.

Recently as XML is becoming the standard of exchanging web documents and public documentations, XML data are increasing in many areas. And the semantic web based on the ontology is appearing for the exact information retrieval. The ontology for not only the text data but also XML data is being required. However, the existing ontology has been constructed manually and it is time and cost consuming. Therefore in this paper, we propose the semi-automatic ontology generation method using the data mining technique, the association rules. Applying the association rules to the XML documents, we intend to find out the conceptual relationships to construct the ontology. Using the conceptual ontology domain level extracted from the XML documents, we construct the ontology by using XML Topic Maps (XTM) automatically.

A logic model for analyzing complex systems’ stability is very useful to many areas of sciences. In the real world, we are enlightened from some natural phenomena such as “biosphere”, “food chain”, “ecological balance” etc. By research and practice, and taking advantage of the orthogonality and symmetry defined by the theory of multilateral matrices, we put forward a logic analysis model of stability of complex systems with three relations, and prove it by means of mathematics. This logic model is usually successful in analyzing stability of a complex system. The structure of the logic model is not only clear and simple, but also can be easily used to research and solve many stability problems of complex systems. As an application, some examples are given.

The octree model, a hierarchical volume description of 3D objects, may be utilized to generate projected images from arbitrary viewing directions, thereby providing an efficient means of the data base for 3D object recognition. The feature points of an occluded object are matched to those of the model object shapes generated automatically from the octree model in viewing directions equally spaced in the 3D coordinates. The best matched viewing direction is calibrated by searching for the 4 pairs of corresponding feature points between the input and the model image projected along the estimated viewing direction. Then the input shape is recognized by matching to the projected shape. Experiment results show good performance of proposed algorithm.

Based on the exponential possibility model, the possibility theoretic clustering algorithm is proposed in this paper. The new algorithm is distinctive in determining an appropriate number of clusters for a given dataset while obtaining a quality clustering result. The proposed algorithm can be easily implemented using an alternative minimization iterative procedure and its parameters can be effectively initialized by the Parzon window technique and Yager’s probability-possibility transformation. Our experimental results demonstrate its success in artificial datasets.

Near Infrared Reflectance (NIR) spectroscopy technique was used to estimate N, P, K and OM content in a loamy mixed soil of Zhejiang, Hangzhou. A total of 165 soil samples were taken from the field, 135 samples spectra were used during the calibration and cross-validation stage. 30 samples spectra were used to predict N, P, K and OM concentration. NIR spectra and constituents were related using Partial Least Square Regression (PLSR) technique. The r between measured and predicted values of N and OM, were 0.925 and 0.933 respectively, demonstrated that NIR spectroscopy have potential to predict accurately this constituents in this soil, not being this way in the prediction of P and K with r, 0.469 and 0.688 respectively, demonstrated a poorly for P and a less successfully for K prediction. The result also shows that NIR could be a good tool to be combined with precision farming application.

A new feature extraction method, called dynamic independent component analysis (DICA), is proposed in this paper. This method is able to extract the major dynamic features from the process, and to find statistically independent components from auto- and cross-correlated inputs. To deal with the regression estimation, we combine DICA with support vector regression (SVR) to construct multi-layer support vector regression. The first layer is feature extraction that has the advantages of robust performance and reduction of analysis complexity. The second layer is the SVR that makes the regression estimation. This kind of soft-sensor estimator was applied to estimation of process compositions in the simulation benchmark of the Tennessee Eastman (TE) plant. The simulation results clearly showed that the estimator by feature extraction using DICA can perform better than that without feature extraction and with other statistical methods for feature extraction.

In recent years, mining with imbalanced data sets receives more and more attentions in both theoretical and practical aspects. This paper introduces the importance of imbalanced data sets and their broad application domains in data mining, and then summarizes the evaluation metrics and the existing methods to evaluate and solve the imbalance problem. Synthetic minority over-sampling technique (SMOTE) is one of the over-sampling methods addressing this problem. Based on SMOTE method, this paper presents two new minority over-sampling methods, borderline-SMOTE1 and borderline-SMOTE2, in which only the minority examples near the borderline are over-sampled. For the minority class, experiments show that our approaches achieve better TP rate and F-value than SMOTE and random over-sampling methods.

In this paper, we present a novel method for real time gesture recognition with 3D Motion History Model (MHM). There are two difficult problems in gesture recognition: the camera view and the duration of gesture. First, we solved the camera view problem which is very difficult in the environment of single directional camera (e.g., monocular or stereo camera). Utilizing 3D-MHM with the disparity information, not only this problem is solved but also the reliability of recognition and the scalability of system are improved. Second, we proposed the dynamic history buffering (DHB) to solve the duration problem that comes from the variation of gesture velocity at every performing time. DHB improves the problem using magnitude of motion. We implemented a real-time system and performed gesture recognition experiments. The system using 3D-MHM achieves better results of recognition than using only 2D motion information.

Recently, the progress of information and communication technologies leads to web-based education. However, one of these problems is that it is very difficult to find out proper educational materials over the billions of unclassified and unrelated web materials. Well-designed directory services and classification systems of educational materials in the Internet are indispensable for effective web-based education. In this paper, we propose novel directory services and classification systems for effective Korean language learning. We analyzed the elementary components of Korean language learning, and exploit them to develop effective directory services and classification systems. We also propose a guideline to develop them. We also consider peer-to-peer networking service as searching and exchanging educational material.

Only a few studies have investigated on how to select component classifiers from a classifier pool. But, the performance of multiple classifier systems depends on the component classifiers as well as the combination methods. A couple of information-theoretic methods selecting the component classifiers by considering the relationship among classifiers are proposed in this paper. These methods are applied to the classifier pool and examine the possible classifier sets for building the multiple classifier systems. A classifier set is selected as a candidate and evaluated with the other classifier sets on the recognition of unconstrained handwritten numerals.

Radial basis function (RBF) networks have been successfully applied to function interpolation and classification problems among others. In this paper, we propose a basis function optimization method using a mixture density model. We generalize the Gaussian radial basis functions to arbitrary covariance matrices, in order to fully utilize the Gaussian probability density function. We also try to achieve a parsimonious network topology by using a systematic procedure. According to experimental results, the proposed method achieved fairly comparable performance with smaller number of hidden layer nodes to the conventional approach in terms of correct classification rates.

We have explored a meta-learning approach to improve the prediction accuracy of a classification system. In the meta-learning approach, a meta-classifier that learns the bias of a classifier is obtained so that it can evaluate the prediction made by the classifier for a given example and thereby improve the overall performance of a classification system. The paper discusses our meta-learning approach in details and presents some empirical results that show the improvement we can achieve with the meta-learning approach in a GA-based inductive learning environment.

Resource-Constrained Project Scheduling Problem(RCPSP) with the Objective of Minimizing Activities’ Cost is one of the critical sub-problems in partner selection of construction supply chain management. In this paper, this type of RCPSP is mathematically modelled firstly. The analysis on the characteristic of the problem shows that the objective function is non-regular and the problem is NP-complete. The basic idea for the solution of the problem is clarified. A genetic algorithm is developed and the parameters of the algorithm are analyzed based on the tests of an example. The proposed GA is demonstrated to be effective based on the results of a computational study on the updated PSPLIB.

A partial rescheduling (PR) heuristic is presented for single machine with unforeseen breakdowns. Unlike a full rescheduling strategy where all unfinished jobs are considered, a PR strategy reschedules partial unfinished jobs which form a PR problem, and shifts the rest jobs to the right according to the solution of the PR problem. The rescheduling problem considers a bi-criterion that optimizes both shop efficiency (i.e. makespan performance of the schedule) and stability (i.e. deviation from the original schedule). A genetic algorithm is developed to solve the PR problem. Extensive computational testing was conducted. The computational results show that the PR heuristic with bi-criterion can significantly improve schedule stability with little sacrifice in efficiency, and provide a reasonable trade-off between solution quality and computational efforts.

An improved genetic algorithm based on information entropy is presented in this paper. A new iteration scheme in conjunction with multi-population genetic strategy, entropy-based searching technique with narrowing down space and the quasi-exact penalty function is developed to solve nonlinear programming problems with equality and inequality constraints. A specific strategy of reserving the most fitness member with evolutionary historic information is effectively used to approximate the solution of the nonlinear programming problems to the global optimization. Numerical examples and an application in molecular docking demonstrate its accuracy and efficiency.

Multi-agent systems form a particular type of distributed artificial intelligence systems. As an important character of players in game, autonomous agents’ learning has become the main direction of researchers. In this paper, based on basic reinforcement learning, multi-agent reinforcement learning with specific context is proposed. The method is applied to RoboCup to learn coordination among agents. In the learning, the game field is divided into different areas, and the action choice is made dependent on the area in which the ball is currently located. This makes the state space and the action space decrease. After learning the optimal joint policy is determined. Comparison experiment between stochastic policy and this optimal policy shows the effectiveness of our approach.

The paper deals with the new algorithm development and comparison of three one-dimensional stock cutting algorithms regarding trim loss. Three possible types of problems used in this study are identified as easy, medium and hard. Approximate method is developed which enables a comparison of solutions of all three types of problems and of the other two stochastic methods. The other two algorithms employed here are Genetic Algorithms (GA) with Improved Bottom-Left (BL) and Simulated Annealing (SA) with Improved BL. Two examples of method implementation for comparison of three algorithms are presented. The approximate method produced the best solutions for easy and medium cutting problems. However, GA works very well in hard problems because of its global search ability.

Genetic algorithm (GA) is an effective method to tackle combinatorial optimization problems. Since the limitation of encoding method, the search space of GA should be regular. Unfortunately, for constraint optimizations, this precondition is unsatisfied. To obtain a regular search space, a commonly used method is penalty functions. But the setting of a good penalty function is difficult. In this paper, a novel algorithm, called search space filling and shrinking algorithm (SSFSA), is proposed. SSFSA first seeks a smaller search space which covers all the feasible domains, then fills the unfeasible search space to acquire a regular search space. Search space shrinking diminishes the search space, so shortens the searching time. Search space filling repairs the irregular search space, and makes GA execute effectively. Experimental results show that SSFSA outperforms penalty methods’.

Intrusion detection has emerged as an important technique for network security. Due to the complex and dynamic properties of intrusion behaviors, machine learning and data mining methods have been widely employed to optimize the performance of intrusion detection systems (IDSs). However, the results of existing work still need to be improved both in accuracy and in computational efficiency. In this paper, a novel reinforcement learning approach is presented for host-based intrusion detection using sequences of system calls. A Markov reward process model is introduced for modeling the behaviors of system call sequences and the intrusion detection problem is converted to predicting the value functions of the Markov reward process. A temporal different learning algorithm using linear basis functions is used for value function prediction so that abnormal temporal behaviors of host processes can be predicted accurately and efficiently. The proposed method has advantages over previous algorithms in that the temporal property of system call data is well captured in a natural and simple way and better intrusion detection performance can be achieved. Experimental results on the MIT system call data illustrate that compared with previous work, the proposed method has better detection accuracy with low training costs.

An architecture is proposed that combines a simple learning method with one of the most natural evaluation systems: imitation controlled by emotions. Using this architecture agents develop behavioral clusters and form a society that improves its ability to reach a given goal over time. Imitation works by observing and applying behavior sequences (episodes). This leads to new and diverse episodes, because the observation introduces small errors. On the other hand, bad episodes are forgotten if they don’t help the agents to satisfy their emotional system that plays the role of an inherent performance measurement. After a while, the agents can be grouped by their typical behavioral patterns. Since these imitated sequences can be seen as “memes” similar to genes in the biological world, this paper explores imitation from the view of memetic proliferation.

We show by simulation that using imitation combined with emotions as evaluation measure tasks can be performed by an agent society without having to specify them in detail. The society’s performance is quantified using an entropy measure to qualitatively evaluate the emerging behavioral clusters.

Canonical genetic algorithms have the defects of pre-maturity and stagnation when applied in optimizing problems. In order to avoid the shortcomings, an adaptive niche hierarchy genetic algorithm (ANHGA) is proposed. The algorithm is based on the adaptive mutation operator and crossover operator to adjust the crossover rate and probability of mutation of each individual, whose mutation values are decided using individual gradient. This approach is applied in Percy and Shubert function optimization. Comparisons of niche genetic algorithm (NGA), hierarchy genetic algorithm (HGA) and ANHGA have been done by establishing a simulation model and the results of mathematics model and actual industrial model show that ANHGA is feasible and efficient in the design of multi-extremum.

Associativity, auto-reversibility and question-answering are the three intrinsic functions to be investigated for the proposed Q’tron Neural Network (NN) model. A Q’tron NN possesses these functions due to its property of local-minima free if it is built as a

known-energy system

which is equipped with the proposed

persistent noise-injection mechanism

. The so-built Q’tron NN, as a result, will settle down if and only if it ‘feels’ feasible, i.e., the energy of its state has been low enough truly. With such a nature, the NN is able to accommodate itself ‘everywhere’ to reach a feasible state autonomously. Three examples, i.e., an associative adder, an

N

-queen solver, and a pattern recognizer are demonstrated in this paper to highlight the concept.

Text categorization has become one of the key techniques for handling and organizing text data. This model is used to classify new article to its most relevant category. In this paper, we propose a novel associative classification algorithm

ACTC

for text categorization.

ACTC

aims at extracting the

k

-best strong correlated positive and negative association rules directly from training set for classification, avoiding to appoint complex support and confidence threshold.

ACTC

integrates the advantages of the previously proposed effective strategies as well as the new strategies presented in this paper. An extensive performance study reveals that the improvement of

ACTC

outperform other rule-based classification approaches on accuracy.

In neural modeling of non-linear dynamic systems, the neural inputs can include any system variable with time delays. To obtain the optimal subset of inputs regarding a performance measure is a combinational problem, and the selection process can be very time-consuming. In this paper, neural input selection is transformed into a model selection problem and a new fast input selection method is used. This method is then applied to the neural modeling of a continuous stirring tank reactor (CSTR) to confirm its effectiveness.

In this paper, we consider a class of time-delay artificial neural networks and obtain practical criteria to test asymptotic stability of the equilibrium of the time-delay artificial neural networks, with or without perturbations. These criteria require verification of the definiteness of a certain matrix, or verification of a certain inequality. Furthermore, we discuss the exponential stability and estimate the exponential convergence rate for time-delay artificial neural networks. The applicability of our results is demonstrated by means of two specific examples

.

The transmission streaming media becomes a challenging study problem for the Web application. The proxy cache for streaming media is an efficient method to solve this problem. In proxy cache, prefix cache is to cache the initial part of the streaming media in the proxy cache, so that there is no startup delay for the inquest of the clients. Segmentation based cache is to cache the length of streaming media, according to the inquest frequency of the clients so as to save the web resources; The proxy cache and the segmentation based cache are both pre-drawing cache methods. In this paper, we proposed the algorithm of high efficient dynamic cache method, based on prefix cache and segmentation based cache strategy. The algorithm carries out real time dynamic cache in the proxy cache and makes the dynamic cache with batch and patching algorithm, transmitted and proxy cached by the server, deal with the requests of more than one clients within a relatively short period of time. Therefore, the web resources used by patching channel and regular channel and release the network burden of the server. Event-driven simulation, introduced to evaluate this algorithm, is very efficient.

There are many features on a palm and different features reflect the different characteristic of a palmprint. Fusion of multiple palmprint features may enhance the performance of palmprint authentication system. In this paper, we investigate the fusion of the textural feature (PalmCode) and the palm-lines. Several fusion strategies have been compared. The experimental results show that the original PalmCode scheme is optimal for the very high security systems, while the fusion of the PalmCode and palm-lines using the Weighted Sum Strategy is the best choice for other systems.

Artificial neural networks have presented their powerful ability and efficiency in nonlinear control, chaotic time series prediction, and many other fields. Reinforcement learning, which is the last learning algorithm by awarding the learner for correct actions, and punishing wrong actions, however, is few reported to nonlinear prediction.

In this paper, we construct a multi-layer neural network and using reinforcement learning, in particular, a learning algorithm called Stochastic Gradient Ascent (SGA) to predict nonlinear time series. The proposed system includes 4 layers: input layer, hidden layer, stochastic parameter layer and output layer. Using stochastic policy, the system optimizes its weights of connections and output value to obtain its prediction ability of nonlinear dynamics. In simulation, we used the Lorenz system, and compared short-term prediction accuracy of our proposed method with classical learning method.