Artificial Intelligence and Soft Computing

In this article, an approximation of the spatiotemporal response of a distributed parameter system (DPS) with the use of the neural network-based principal component analysis (PCA) is considered. The presented approach is carried out using two different neural structures: single-layer network with unsupervised, generalized Hebbian learning (GHA-PCA) and two-layer feedforward network with supervised learning (FF-PCA). In each case considered, the effect of the number of units in the network projection layer on the mean square approximation error (MSAE) and on the data compression ratio is analysed.

This paper presents the parallel architecture of the Recurrent Multi Layer Perceptron learning algorithm. The proposed solution is based on the high parallel three dimensional structure to speed up learning performance. Detailed parallel neural network structures are explicitly shown.

The investigation of solar-like oscillations for probing the star interiors has encountered a tremendous growth in the last decade. For ground based observations the most important difficulties in properly identifying the true oscillation frequencies of the stars are produced by the gaps in the observation time-series and the presence of atmospheric plus the intrinsic stellar granulation noise, unavoidable also in the case of space observations. In this paper an innovative neuro-wavelet method for the reconstruction of missing data from photometric signals is presented. The prediction of missing data was done by using a composite neuro-wavelet reconstruction system composed by two neural networks separately trained. The combination of these two neural networks obtains a ”forward and backward” reconstruction. This technique was able to provide reconstructed data with an error greatly lower than the absolute a priori measurement error. The reconstructed signal frequency spectrum matched the expected spectrum with high accuracy.

Training feed-forward neural networks can take a long time when there is a large amount of data to be used, even when training with more efficient algorithms like Levenberg-Marquardt. Parallel architectures have been a common solution in the area of high performance computing, since the technology used in current processors is reaching the limits of speed. An architecture that has been gaining popularity is the GPGPU (General-Purpose computing on Graphics Processing Units), which has received large investments from companies such as NVIDIA that introduced CUDA (Compute Unified Device Architecture) technology. This paper proposes a faster implementation of neural networks training with Levenberg-Marquardt algorithm using CUDA. The results obtained demonstrate that the whole training time can be almost 30 times shorter than code using Intel Math Library (MKL). A case study for classifying electrical company customers is presented.

Sufficient conditions for uniform convergence of general regression neural networks, based on the orthogonal series-type kernel, are given. The convergence is guarantee even if variance of noise diverges to infinity. Simulation results are presented.

Strong convergence of general regression neural networks is proved assuming non-stationary noise. The network is based on the orthogonal series-type kernel. Simulation results are discussed in details.

Sufficient conditions for strong convergence of recursive general regression neural networks are given assuming nonstationary noise. The orthogonal series-type kernel is applied. Simulation results show convergence even if variance of noise diverges to infinity.

In this paper we developed a new architecture of neural networks for generating nomograms based on series of data vectors. The paper was inspired by the XIII Hilbert’s problem which was presented 1900 in the context of nomography, for the particular nomographic construction. The problem was solved by V. Arnold (a student of Andrey Kolomogorov) in 1957. For numeric data of unknown functional relation we developed the

incidental neural networks

as nomograms generators – the graphic calculating devices.

The paper presents some novel methods of the activation function selection in the last hidden layer of a multilayer perceptron. For this selection, the least squares method is used. The proposed ways make it possible to decrease the cost function value. They enable achievement of a good compromise between the network complexity and the results being obtained. The methods do not require a start of learning of neural networks from the very beginning. They fit very well for improvement of the action of learnt multilayer perceptrons. They may be particularly useful for construction of the devices under microprocessor control, that have not a big memory nor computing power.

Today, majority of collected data and information are usually passively stored in data bases and in various kinds of memory cells and storage media that let them do nothing more than waiting for being used by some algorithms that will read, write or modify them. Nowadays, the majority of computational techniques do not allow pieces of information to associate with each other automatically. This paper introduces a novelty theory that lets information be free and active. There is allowed that some pieces of information can automatically and autonomously associate with the other pieces of it after some introduced associative rules characteristic also for biological information systems. As a result of this, each new information has an automatic impact on information processing in a brainlike artificial neural structure that can enable machines to associate various pieces of information automatically and autonomously. It can also enable machines actively perform some cognitive and thinking processes and constitute real artificial intelligence in the future.

A problem of learning in non-stationary environment is solved by making use of order statistics in combination with the Parzen kernel-type regression neural network. Probabilistic properties of the algorithm are investigated and weak convergence is established. Experimental results are presented.

This paper presents the recursive least squares method, combined with the general regression neural networks, applied to solve the problem of learning in time-varying environment. The general regression neural network is based on the orthogonal-type kernel functions. The appropriate algorithm is presented in a recursive form. Sufficient simulations confirm empirically the convergence of the algorithm.

A number of researchers headed by E. Gardner have proved that a maximum achievable memory load of binary perceptron is 2. A learning algorithm allowing reaching and even exceeding the critical load was proposed. The algorithm was reduced to solving the linear programming problem. The proposed algorithm is sequel to Krauth and Mezard ideas. The algorithm makes it possible to construct networks storage capacity and noise stability of which are comparable to those of Krauth and Mezard algorithm. However suggested modification of the algorithm outperforms.

In the present thesis the author undertakes the problem of the objects selecting on pictures. The novel conception of using depth map as a base to objects marking was proposed here. objects separation can be done on the base of depth (disparity), corresponding to points that should be marked. This allows for elimination of textures, occurring in background and also on objects. The object selection process must be preceded by picture’s depth analysis. This can be done by the novel neural structure: Self-Correcting Neural Network. This structure is working point-by-point with no picture’s segmentation before.

The direct application of a neural model in Model Predictive Control (MPC) algorithms results in a nonlinear, in general non-convex, optimisation problem which must be solved on-line. A linear approximation of the model for the current operating point can be used for prediction in MPC, but for significantly nonlinear processes control accuracy may be not sufficient. MPC algorithm in which the neural model is linearised on-line along a trajectory is discussed. The control policy is calculated from a quadratic programming problem, nonlinear optimisation is not necessary. Accuracy and computational burden of the algorithm are demonstrated for a high-purity high-pressure distillation column.

The paper presents basic notions of web mining, radial basis function (RBF) neural networks and

ε

-insensitive support vector machine regression (

ε

-SVR) for the prediction of a short time series (website of the University of Pardubice, Czech Republic). There are various short time series according to different visitors or interest of visitors (students, employees, documents). Further, a model (including RBF neural networks and

ε

-SVRs) was developed for short time series prediction. The model includes decomposition of data to training and testing data set using the cluster procedure. The next part of the paper describes the predictions of the web domain visits, which depend on this model, as well as outlines an analysis of the results.

In this work we study a simplified model of a neural activity flow in networks, whose connectivity is based on geometrical embedding, rather than being lattices or fully connected graphs. We present numerical results showing that as the spectrum (set of eigenvalues of adjacency matrix) of the resulting activity-based network develops a scale-free dependency. Moreover it strengthens and becomes valid for a wider segment along with the simulation progress, which implies a highly organised structure of the analysed graph.

In this paper we study probabilistic neural networks based on the Parzen kernels. Weak convergence is established assuming time-varying noise. Simulation results are discussed in details.

A recursive version of the Parzen-type general regression neural network is studied. Strong convergence is established assuming time-varying noise. Experimental results are discussed in details.

Self-Organising Maps (SOM) are Artificial Neural Networks used in Pattern Recognition tasks. Their major advantage over other architectures is human readability of a model. However, they often gain poorer accuracy. Mostly used metric in SOM is the Euclidean distance, which is not the best approach to some problems. In this paper, we study an impact of the metric change on the SOM’s performance in classification problems. In order to change the metric of the SOM we applied a distance metric learning method, so-called ’Large Margin Nearest Neighbour’. It computes the Mahalanobis matrix, which assures small distance between nearest neighbour points from the same class and separation of points belonging to different classes by large margin. Results are presented on several real data sets, containing for example recognition of written digits, spoken letters or faces.

In the on-line data processing it is important to detect a novelty as soon as it appears, because it may be a consequence of gross errors or sudden change in the analysed system. In this paper we present a framework of novelty detection, based on the robust neural network. To detect novel patterns we compare responses of two autoregressive neural networks. One of them is trained with a robust learning algorithm designed to remove the influence of outliers, while the other uses simple training, based on the least squares error criterion. We present also a simple and easy to use approach that adapts this technique to data streams. Experiments conducted on data containing novelty and outliers have shown promising performance of the new method, applied to analyse temporal sequences.

A method for solving both, ordinary and partial, non-linear differential equations (DE) by means of the feed-forward artificial neural networks (ANN) is presented in this paper. Proposed approach consist in training ANN in such a way, that it approximates a function being a particular solution of DE and all its derivatives, up to the order of the equation. This is achieved by special construction of the cost function which contains informations about derivatives of the network. ANNs with sigmoidal activation functions in hidden nodes, thus infinitely differentiable, are considered in this paper. Illustrative examples of the solution of a non-linear DE are also presented.

The article is dedicated to the possibilities of practical application of artificial neural networks in designing parameters of steel vacuum carburization processes and preparing for cooling in high-pressure gas. In the following sections, the nature of vacuum carburization technology, the course of research on the precipitation phenomena, the construction of an artificial neural network and the algorithm of searching process parameters have been presented.

In this paper, a new method for dealing with an unbalanced linguistic term set is introduced. The proposed method is a modification of the 2-tuple linguistic model, in which we use a set of extended linguistic terms. The extended linguistic term is a pair that consists a linguistic label and a value of correction factor which describes the term shift relative to its position in an equidistant term set. This modification allows us to obtain the method that is computationally less expensive and give simpler semantics than method based on linguistic hierarchies.

In the paper a new method of fuzzy clustering basing on fuzzy features is presented. Objects are described by set of features with intutionistic fuzzy values. Generally, the method uses the concept of modified fuzzy c-means procedure applied to intuitionistic fuzzy data which describes the features. New distance measure between data and cluster centers is suggested. Some examples of clustering results are presented. The method is efficient and very fast.

This paper presents a new approach to comparing interval-valued intuitionistic fuzzy values. The interval score and accuracy functions are used to build the “net profit” and “risk” local criteria. These criteria are aggregated in a generalized criterion taking into account their weights, which depend on the risk aversion of a decision maker. As opposed to the known methods, a new approach makes it possible to estimate the strength of relations between interval-valued intuitionistic fuzzy values. Using some numerical examples, it is shown that the proposed approach provides intuitively clear results.

This paper presents a critical analysis of conventional operations on intuitionistic fuzzy values (

IFVs

) and their applicability to the solution of multiple criteria decision making (

MCDM

) problems in the intuitionistic fuzzy setting. A set of operations on

IFVs

based on the interpretation of intuitionistic fuzzy sets in the framework of the Dempster-Shafer theory of evidence (

DST

) is proposed and analyzed. This interpretation makes it possible to represent mathematical operations on

IFVs

as operations on belief intervals. The corresponding method for aggregation of local criteria presented by

IFVs

in the framework of

DST

is proposed and analyzed. The proposed approach allows us to solve

MCDM

problems without intermediate defuzzification when not only criteria, but their weights are

IFVs

. The advantages of the proposed approach are illustrated by numerical examples.

Linear PID algorithms commonly used in industry might perform insufficiently when controlling nonlinear operating systems. Solutions such as the fuzzy PID controller can exchange the linear PID controller because it develops a nonlinear control surface. The main advantage of the fuzzy PID controller is the ability to adjust to a controlled plant by the rule based modification, nonlinear membership function application and inference rule selection. However, the tuning process is one of the most difficult steps in the fuzzy PID controller designing and therefore discourages most practical applications.

A simplification of the fuzzy PID controller tuning process is described in this article. The presented methodology allows fast transformation from a classic PID algorithm into a fuzzy PID algorithm. A proposed algorithm is tested on a programmable PLC which is a typical industrial implementation platform. A temperature stabilization is chosen as the controlled plant and some experimental results are then described. In conclusion authors suggest directions for similar real time fuzzy PID algorithm implementations.

Ordered fuzzy numbers (OFN) as generalization of convex fuzzy numbers represented in parametric form and invented by the second and the third authors and their coworker in 2002, make possible to utilize the fuzzy arithmetic and to construct the lattice structure on them. Fuzzy inference mechanism and implications are proposed together with step fuzzy numbers that may be used for approximations as well as for constructing new fuzzy sets of type two.

In this paper we propose a new approach to combine unsupervised and supervised vector quantization for clustering and fuzzy classification using the framework of neural vector quantizers like self-organizing maps or neural gas. For this purpose the original cost functions are modified in such a way that both aspects, unsupervised vector quantization and supervised classification, are incorporated. The theoretical justification of the convergence of the new algorithm is given by an adequate redefinition of the underlying dissimilarity measure now interpreted as a dissimilarity in the data space combined with the class label space. This allows a gradient descent learning as known for the original algorithms. Thus a semi-supervised learning scheme is achieved. We apply this method for a spectra image cube of remote sensing data for landtype classification. The obtained fuzzy class visualizations allow a better understanding and interpretation of the spectra.

Falls are major causes of mortality and morbidity in the elderly. However, prevalent methods only utilize accelerometers or both accelerometers and gyroscopes to separate falls from activities of daily living. This makes it not easy to distinguish real falls from fall-like activities. The existing CCD-camera based solutions require time for installation, camera calibration and are not generally cheap. In this paper we show how to achieve reliable fall detection. The detection is done by a fuzzy inference system using low-cost Kinect and a device consisting of an accelerometer and a gyroscope. The experimental results indicate high accuracy of the detection and effectiveness of the system.

Defuzzification operators, that play the main role when dealing with fuzzy controllers and fuzzy inference systems, are discussed for convex as well for ordered fuzzy numbers. Three characteristic conditions are formulated for them. It is shown that most of known defuzzification functionals meet these requirements. Some approximation methods for determining of the functionals are given and then applied.

The ordered weighted averaging (OWA) operator uses the weights assigned to the ordered values rather than to the specific criteria. This allows one to model various aggregation preferences, preserving simultaneously the impartiality (neutrality) with respect to the individual attributes. The determination of ordered weighted averaging (OWA) operator weights is a crucial issue of applying the OWA operator for decision making. This paper considers determining monotonic weights of the OWA operator by minimization the mean absolute deviation inequality measure. This leads to a linear programming model which can also be solved analytically.

Efficient Model Predictive Control (MPC) algorithms based on fuzzy Wiener models with advanced methods of prediction are proposed in the paper. The methods of prediction use values of future control changes which were derived by the MPC algorithm in the last iteration. Such an approach results in excellent control performance offered by the proposed algorithms. Moreover, they are formulated as numerically efficient quadratic optimization problems. Advantages of the proposed fuzzy MPC algorithms are demonstrated in the control systems of a nonlinear plant.

Low physical activity (PA), and often concomitant overweight in the developmental age are well documented risk factors for cardiovascular diseases (CVD) and many other chronic civilization diseases. Regular monitoring of health related physical fitness (H-RF) is an important part of early prevention and school health education. An assessment of components of H-RF is complex and controversial. In the assessment of H-RF components, systems of fuzzy inference based on simple linguistic variables can be used. The paper presents a system intended to support the evaluation of the H-RF components based on the EUROFIT battery tests and the anthropometric measurements. A basis of the system is the EUROFIT calculator which converts absolute results of individual trials to standarized values and the fuzzy inference system for four H-RF components (Morphological, Cardiorespiratory, Musculoskeletal and Motor Fitness). The system is implemented in MS Visual Studio in C# and has a friendly graphical interface for archiving test results. An application of fuzzy inference elements in the evaluation of the H-RF components is a new approach that can be used in monitoring and rational planning of PA dosing in prophylaxis and therapy.

This paper presents the fuzzy regression approach to the automotive industry optimization problem. The flywheel assembly process is subject to investigation, as its parameters require optimization. The paper contains: problem definition, presentation of the measured data and the final analysis with two alternative approaches: the fuzzy regression and the classical regression. The benefits of the fuzzy regression approach are shown in the case of small size samples.

Along with technological developments we observe an increasing amount of stored and processed data. It is not possible to store all incoming data and analyze it on the fly. Therefore many researchers are working on new algorithms for data stream mining. New algorithm should be fast and should use a small amount of memory. We will consider the problem of data stream classification. To increase the accuracy we propose to use an ensemble of classifiers based on a modified FID3 algorithm. The experimental results show that this algorithm is fast and accurate. Therefore it is adequate tool for data stream classification.

Metaset is a new concept of set with partial membership relation. It is designed to represent and process vague, imprecise data – similarly to fuzzy sets. Metasets are based on the classical set theory primitive notions. At the same time they are directed towards efficient computer implementations and applications. The degrees of membership for metasets are expressed as finite binary sequences, they form a Boolean algebra and they may be evaluated as real numbers too. Besides partial membership, equality and their negations, metasets allow for expressing a hesitancy degree of membership – similarly to intuitionistic fuzzy sets. The algebraic operations for metasets satisfy axioms of Boolean algebra.

This paper is a continuation of our previous works on correlation coefficients of Atanassov’s intuitionistic fuzzy sets (A-IFSs). The Pearson’s coefficient we discuss here yields the strength of relationship between the A-IFSs and also indicates the direction of correlation (positive or negative). The proposed correlation coefficient takes into account all three terms describing an A-IFS (membership values, non-membership values, and the hesitation margins).

In many applications, spatial data is often prone to uncertainty and imprecision. To model this, fuzzy regions have been developed. Our initial model was a fuzzy set over a two dimensional domain, allowing for fuzzy regions and fuzzy points to be modelled. The model had some limitations: all points where treated independently, and it was not possible to group points together. Furthermore, the model depended on meta-information to specify the interpretation. The model was extended to a level-2 fuzzy region to overcome these limitations; here the calculation and interpretation of the surface area will be considered.

In this paper we propose the combination of fuzzy c-means for clustering with neighborhood cooperativeness from the neural gas vector quantizer. The new approach avoids the sensitivity of fuzzy c-means with respect to initialization as it is known from neural gas compared to crisp c-means. Thereby, the neural gas paradigm of neighborhood offers a greater flexibility than those of the self-organizing map, which was combined with fuzzy c-means before. However, a careful reformulation of neighborhood has to be done to keep the validity of the convergence proof of this previous approach. We demonstrate the properties for an artificial as well as for real world data.

The new epoch-incremental reinforcement learning algorithm with fuzzy approximation of action-value function is developed. This algorithm is practically tested in the control of the mobile robot which realizes goal seeking behavior. The obtained results are compared with results of fuzzy version of reinforcement learning algorithms, such as Q(0)-learning, Q(

λ

)-learning, Dyna-learning and prioritized sweeping. The adaptation of the fuzzy approximator to the model based reinforcement learning algorithms is also proposed.

In this paper a new family of kernel functions for SVM classifiers, based on a statistically–induced measure of distance between observations in the pattern space, is proposed and experimentally evaluated in the context of binary classification problems. The application of the proposed approach improves the accuracy of results compared to the case of training without postulated enhancements.

Numerical results outperform those of the SVM with Gaussian and Laplace kernels.

We consider a problem of selection of parameters in a classifier based on the average of kernel density estimators where each estimator corresponds to a different data “resolution”. The selection is based on adjusting parameters of the estimators to minimize a substitute of the misclassification ratio. We experimentally compare the misclassification ratio and parameters selected for benchmark data sets by the introduced algorithm with these values of the algorithm’s baseline version. In order to place the classification results in a wider context, we compare them with results of other popular classifiers.

In the paper the ensemble of dipolar trees for analysis of competing risks is proposed. The tool is build on the base of the learning sets, which contain the data from clinical studies following patients response for a given treatment. In case of competing risks many types of response are investigated. The proposed method is able to cope with incomplete (censored) observations and as a result, for a given set of covariates and a type of event, returns the aggregated cumulative incidence function.

An open source framework for general gesture recognition is presented and tested with isolated signs of sign language. Other than common systems for sign language recognition, this framework makes use of Kinect, a depth camera which makes real-time 3D-reconstruction easily applicable. Recognition is done using hidden Markov models with a continuous observation density. The framework also offers an easy way of initializing and training new gestures or signs by performing them several times in front of the camera. First results with a recognition rate of 97% show that depth cameras are well-suited for sign language recognition.

A few years ago a new classifier ensemble method, called rotation forest, was devised. The technique applies Principal Component Analysis to rotate the original feature axes in order to obtain different training sets for learning base classifiers. In the paper we report the results of the investigation aimed to compare the predictive performance of rotation forest with random forest models, bagging ensembles and single models using two popular algorithms M5 tree and multilayer perceptron. All tests were carried out in the WEKA data mining system within the framework of 10-fold cross-validation and repeated holdout splits. A real-world dataset of sales/purchase transactions derived from a cadastral system served as basis for benchmarking the methods.

Support Vector Machines (SVM’s) with various kernels have become very successful in pattern classification and regression. However, single kernels do not lead to optimal data models. Replacing the input space by a kernel-based feature space in which the linear discrimination problem with margin maximization is solved is a general method that allows for mixing various kernels and adding new types of features. We show here how to generate locally optimized kernels that facilitate multi-resolution and can handle complex data distributions using simpler models than the standard data formulation may provide.

We present a practical application of Hierarchical Classifier with overlapping clusters to the problem of finding the minimal synchronizing word length of a given finite automaton. We compare our approach with a single neural network model. Using a certain representation of automaton as the classifier’s input we improve HC efficiency and we are able to analyze the relation between particular automata features and minimal synchronizing lengths.

Our aim is to propose a new look at the dimensionality reduction in pattern recognition problems by extracting part of variables that are further called external context variables. We show how to incorporate them into the Bayes classification scheme with loss functions that depend on class labels that are ordered. Then, the general form of the optimal context sensitive classifier is derived and the learning method that is based on kernel approximation is proposed.

The SVM algorithm is one of the most frequently used methods for the classification process. For many domains, where the classification problems have many features as well as numerous instances, classification is a difficult and time-consuming task. For this reason, the following paper presents the CSR-GPU-SVM algorithm which accelerates SVM training for large and sparse problems with the use of the CUDA technology. Implementation is based on the SMO (Sequential Minimal Optimization) algorithm and utilizes the CSR(Compressed Sparse Row) sparse matrix format. The proposed solution allows us to perform efficient classification of big datasets, for example rcv1 and newsgroup20, for which classification with dense representation is not possible. The performed experiments have proven the accelerations in the order of 6 - 35 training times compared to original LibSVM implementation.

Nonnegative Matrix Factorization (NMF) is an emerging unsupervised learning technique that has already found many applications in machine learning and multivariate nonnegative data processing. NMF problems are usually solved with an alternating minimization of a given cost function, which leads to non-convex optimization. For this approach, an initialization for the factors to be estimated plays an essential role, not only for a fast convergence rate but also for selection of the desired local minima. If the observations are modeled by the exact factorization model (consistent data), NMF can be easily obtained by finding vertices of the convex polytope determined by the observed data projected on the probability simplex. For an inconsistent case, this model can be relaxed by approximating mean localizations of the vertices. In this paper, we discuss these issues and propose the initialization algorithm based on the analysis of a geometrical structure of the observed data. This approach is demonstrated to be robust, even for moderately noisy data.

The paper contains test of corner detectors applied in finding characteristic points on 3D revolving objects. Five different algorithm are presented starting from historical Moravec detector and ending at newest ones, such as SUSAN and Trajkovic.

Since the algorithms are compared from the perspective of use for 3D modeling, the count of detected points and their localization is compared. The modeling process uses a series of photos and requires finding a projection of 3D point to two or three subsequent photos. The quality of algorithms is discussed on the base of the ability to detect modeled objects’ corners and immunity to noise. The last researched aspect is the computation cost.

The presented tests show that the best results are given by Shi–Tomasi operator. The detector does find false corners on noisy images, thus SUSAN operator may be used instead.

In this paper, we propose a hierarchical action recognition system applying Fisher discrimination dictionary learning via sparse representation classifier. Feature vectors used to represent certain actions are first generated by employing local features extracted from motion field maps. Sparse representation classification (SRC) are then employed on those feature vectors, in which a structured dictionary for classification is learned applying Fisher discrimination dictionary learning (FDDL). We tested our algorithms on Weizmann human database and KTH human database, and compared the recognition rates with other modeling methods such as k-nearest neighbor. Results showed that the action recognition system applying FDDL can achieve better performance despite that the learning stage for the Fisher discrimination dictionary can converge within only several iterations.

Human-Computer Interaction (HCI) is one of the most rapidly developing fields of computer applications. One of approaches to HCI is based on gestures which are in many cases more natural and effective than conventional inputs. In the paper the problem of gesture recognition is investigated. The gestures are gathered from the dedicated motion capture system, and further evaluated by 3 different preprocessing procedures and 4 different classifier. Our results suggest that most of the combinations produce adequate recognition rate, with appropriate signal normalization being the key element.

The active appearance model (AAM) has proven to be a powerful tool for modeling deformable visual objects. AAMs are nonlinear parametric models in terms of the relation between the pixel intensities and the parameters of the model. In this paper, we propose a fitting procedure for a 3D AAM based on kernel methods for regression. The use of kernel functions provides a powerful way of detecting nonlinear relations using linear algorithms in an appropriate feature space. For analysis, we have chosen the relevance vector machines (RVM) and the kernel ridge method. The statistics computed on data generated with our 3D AAM implementation show that the kernel methods give better results compared to the linear regression models. Although they are less computational efficient, due to their higher accuracy the kernel methods have the advantage of reducing the searching space for the 3D AAM fitting algorithm.

In this paper an analytical iterative approach to the problem of image reconstruction from parallel projections is presented. The reconstruction process is performed using Expectation Minimization algorithm. Experimental results show that the appropriately designed reconstruction procedure is able to reconstruct an image with better quality than obtained using the traditional convolution/ back-projection algorithm.

The aim of this paper is to present a new method of two-dimensional shape recognition. The method is based on dependence vectors which are fractal features extracted from the partitioned iterated function system. The dependence vectors show the dependency between range blocks used in the fractal compression. The effectiveness of our method is shown on four test databases. The first database was created by the authors and the other ones are: MPEG7 CE-Shape-1PartB, Kimia-99, Kimia-216. Obtained results have shown that the proposed method is better than the other fractal recognition methods of two-dimensional shapes.

Recently, the field of CBIR has attracted a lot of attention in the literature. In this paper, the problem of visually similar image retrieval has been investigated. For this task we use the methods derived from the Bag of Visual Words approach, such as Scale Invariant Feature Transform (SIFT) for identifying image keypoints and K-means to build a visual dictionary. To create a ranking of similar images, a novel Ranking by K-means Voting algorithm is proposed. The experimental section shows that our method works well for similar image retrieval. It turned out that our results are more accurate in comparison with a classical similarity measure based on the Euclidean metric in the order of 6% - 15%.

A method of Hurwitz-Radon Matrices (MHR) is proposed to be used in parametrization and interpolation of contours in the plane. Suitable parametrization leads to curvature calculations. Points with local maximum curvature are treated as feature points in object recognition and image analysis. The matrices are skew-symmetric and possess columns composed of orthogonal vectors. The operator of Hurwitz-Radon (OHR), built from these matrices, is described. It is shown how to create the orthogonal OHR and how to use it in a process of contour parametrization and curvature calculation.

In this paper, a new, glove-free method for recognition of fingerspelled acronyms using hierarchical temporal memory has been proposed. The task is challenging because many signs look similar from the camera viewpoint. Moreover handshapes are distorted strongly as a result of coarticulation and motion blur, especially in the fluent fingerspelling. In the described work, the problem has been tackled by applying the new, bio-inspired recognition engine, based on structural and functional properties of mammalian neocortex, robust to local changes shape descriptors, and a training scheme allowing for capture possible handshape deformations in a manner that is lexicon independent.

This paper proposes a method of tracking the lips in the system of audio-visual speech recognition. Presented methods consists of a face detector, face tracker, lip detector, lip tracker, and word classifier. In speech recognition systems, the audio signal is exposed to a large amount of acoustic noise, therefor scientists are looking for ways to reduce audio interference on recognition results. Visual speech is one of the sources that is not perturbed by the acoustic environment and noise. To analyze the video speech one has to develop a method of lip tracking. This work presents a method for automatic detection of the outer edges of the lips, which was used to identify individual words in audio-visual speech recognition. Additionally the paper also shows how to use video speech to divide the audio signal into phonemes.

A common task in the field of document digitization for information retrieval is separating text and non-text elements. In this paper an innovative approach of recognizing patterns is presented. Statistical and structural features in arbitrary number are combined into a rating tree, which is an adapted decision tree. Such a tree is trained for character patterns to distinguish text elements from non-text elements. First experiments in a binarization application have shown promising results in significant reduction of false-positives without producing false-negatives.

This paper describes a novel image retrieval method for parasite detection based on the analysis of digital images captured by the camera from a microscope. In our approach we use several image processing methods to find known parasite shapes. At first, we use an edge detection method with edge representation by vectors. The next step consists in clustering edges fragments by their normal vectors and positions. Then grouped edges fragments are used to perform elliptical or circular shapes fitting as they resemble most parasite forms. This approach is invariant from rotation of parasites eggs or the analyzed sample. It is also invariant to scale of digital images and it is robust to overlapping shapes of parasites eggs thanks to the ability to reconstructing elliptical or other symmetric shapes that represent the eggs of parasites. With this solution we can also reconstruct incomplete shape of parasite egg which can be visible only in some part of the retrieved image.

The following paper describes a novel lipreading procedure based on dynamic programming. We proposed a new method of outer lip contour extraction and representation. Lip shapes, corresponding to selected group of visems, are firstly extracted using dynamic programming and then approximated by B-splines. Coordinates of B-spline control points form final feature vector used for visem recognition task. The discontinuity of lip gradient image is addressed by dynamic programming technique. This has the advantage of global minimum detection and consequently optimal lip contour extraction. Experiments for Polish language utterances show that seven classes of visems can be recognized with 75% accuracy.

The goal of this paper is to present data structures in 3D face recognition systems emphasizing the role of meshes and depth maps. 3D face recognition systems are still in development since they use different data structures. There is no standarized form of 3D face data. Dedicated hardware (3D scanners) usually provide depth maps of objects, which is not sufficiently flexible data sturcture. Meshes are huge structures and operating on them is difficult and requieres a lot of resources. In this paper, we present advantages and disadvantages of both types of data structures in 3d face recognition systems.

This paper presents a system for facial expression recognition which is designed to detect spontaneous emotions. The goal was to detect human aggression. Using a face detection algorithm, a representation of the human face was created. Then, the face texture was encoded with Gabor filter and Local Binary Pattern (LBP) operator. These techniques were used to find the feature set in emotion recognition. As a classifier, a Support Vector Machine (SVM) was applied. The system constructed was tested with spontaneous emotions for aggression detection. The numerical results indicate that the presented classifier achieved an 85% correctness recognition coefficient.

A new set of attributes combining color and SURF-based histograms coupled with a SVM classifier to enhance visual based autonomous aerial navigation is proposed. These new features are used for region classification with aerial images in order to speed up the UAV (Unmanned Aerial Vehicles) localization performed by image matching using only reference images according to the region classification. Experimental results comparing the proposal with color or SURF only attributes are presented. In the experiments the UAV localization task can be performed four times faster using the proposed approach, however the performance gain can be still bigger for large datasets of reference images.

In the world of computer imaging, we still do not have a good and fast enough method for image searching. This is because science is still not able to imitate fully functions of the human brain. When humans think about images, they do not think about mathematical formulas, matrices, histograms etc. Those mathematical and algorithmic methods are very good for e.g. computer face detection or number plate recognition, but we cannot directly use them for analyzing a whole image and for searching in a set of thousands or even millions of images. On the other hand, computers are able to scan millions of documents, searching for some phrase or even a single word. Fast text search is fully supported by a majority of significant database systems such as Oracle, PostgreSQL or MS SQL Server. The paper presents fast text search engine from another point of view, that is, its application in content based image retrieval.

The estimation of full body pose in monocular images is a very difficult problem. In 3D-model based motion tracking the challenges arise as at least one-third of degrees of freedom of the human pose that needs to be recovered is nearly unobservable in any given monocular image. In this paper, we deal with high dimensionality of the search space through estimating the pose in a hierarchical manner using Particle Swarm Optimization. Our method fits the projected body parts of an articulated model to detected body parts at color images with support of edge distance transform. The algorithm was evaluated quantitatively through the use of the motion capture data as ground truth.

This paper presents the design and application of novel device for real time traffic sign detection and recognition on a hardware platform powered by Intel® Atom

TM

processor. Image frames from standard and relatively cheap web cameras are processed using OpenCV library [7][2]. An innovative method is proposed for traffic sign detection phase. Two color models are used for image segmentation and detection of traffic sign. Many well-known and described tactics have been tested and rated. Implemented in OpenCV Library functions for pattern recognition method are also used in main algorithm. Experimental results of traffic sign detection and recognition are described. The prototype was implemented as part of the Master Thesis at Cracow University of Technology [1].

A real-time object tracking algorithm is presented based on the on-line support vector machine (SVM) scheme. A new training framework is proposed, which enables us to select reliable training samples from the image sequence for tracking. Multiple candidate regeneration, a statistical method, is employed to decrease the computational cost, and a directional-edge-based feature representation algorithm is used to represent images robustly as well as compactly. The structure of the algorithm is designed especially for real-time performance, which can extend the advantages of SVM to most of the general tracking applications. The algorithm has been evaluated on challenging video sequences and showed robust tracking ability with accurate tracking results. The hardware implementation is also discussed, while verification has been done to prove the real-time ability of this algorithm.

In the Social Web, folksonomies and other similar knowledge organization techniques may suffer limitations due to both different users’ tagging behaviours and semantic heterogeneity. In order to estimate how a social tagging network organizes its resources, focusing on sharing (implicit) conceptual schemes, we apply an agent-based reconciliation knowledge system based on Formal Concept Analysis. This article describes various experiments that focus on conceptual structures of the reconciliation process as applied to Delicious bookmarking service. Results will show the prevalence of sharing tagged resources in order to be used by other users as recommendations.

This works contains a proposition for a modern Web service description, where functionality of Web service operations is defined with a set of federated Local Controlled Vocabularies (LCV). The LCVs serve as a referral platform for functionality definition with a phrase schema. This schema allows for describing every Web service operation in terms of main action associated with some object extended with an arbitrary number of supplements and marked with desired non functional properties. The proposed description argues for federated LCV instead of centralised fully fledged ontology based effort due to the cost, scalability and performance issues simultaneously maintaining the high level of expressivity unreachable for standard Information Retrieval systems used in Web service retrieval. This work concludes in presentation of mechanism that allows for query matching on envisioned structure along with experiment results and discussion on possible enhancements.

Development of advanced systems requires new methods to improve quality and efficiency of engineering processes, and to assist management of complex models encompassing different engineering disciplines. Methods such as model-driven development and domain-specific modeling facilitate development from this perspective but reduce interoperability and other prospects of rationalizing processes, on the other hand. An approach applying OWL semantics and reasoning to models is presented with examples to support industrial control application engineering. Using the methods, generalized classifications are inferred from instance models and combined with generic engineering knowledge maintained in ontologies. Reasoning allows identifying assemblies and structures outside the scope of traditional modeling to detect flaws and error-prone designs. The results indicate that OWL semantics and reasoning can be used as a supplement furthering typical development practices.

The collective classification problem for big data sets using MapReduce programming model was considered in the paper. We introduced a proposal for implementation of label propagation algorithm in the network. The method was examined on real dataset in telecommunication domain. The results indicated that it can be used to classify nodes in order to propose new offerings or tariffs to customers.

In this paper, we propose a novel KMS by using semantic wiki framework based on a centralized Global Wiki Ontology (GWO). The main aim of this system is

i

) to collect as many organizational resources as possible, and

ii

) to maintain semantic consistency of the system. During enriching the KMS in a particular domain, not only linguistic resources but also conceptual structures can be efficiently captured from multiple users, and more importantly, the resources can be automatically integrated with the GWO of the KMS in the real time. Once users add new organization resources, the proposed KMS can formalize and contextualize them into a set of triplets by referring to a predefined pattern-triplet mapping table and the GWO. Especially, since the ontology matcher is interleaved, the KMS can determine whether the new resources are semantically conflicted with the GWO.

Modeling users’ information interests and needs is one of the most important tasks in the area of personalization in information retrieval domain. In this paper the statistical model of information retrieval system is considered. A method for tuning the user profile based on analysis of user preferences dynamics is experimentally evaluated to check whether with growing history of user activity the created user profile can come closer to his preferences. As statistical analysis of series of simulations have shown, proposed method of user profile actualization is effective in the sense of distance between user preferences and his profile.

The importance of research on knowledge management is growing due to recent issues with big data. The most fundamental steps in knowledge management are the extraction and construction of terminologies. Terms are often expressed in various forms and the term variations play a negative role, becoming an obstacle which causes knowledge systems to extract unnecessary knowledge. To solve the problem, we propose a method of term normalization which finds a normalized form (original and standard form defined in dictionaries) of variant terms. The method employs a couple of characteristics of terms: one is appearance similarity, which measures how similar terms are, and the other is context similarity which measures how many clue words they share. Through experiment, we show its positive influence of both similarities in the term normalization.

The cascaded fuzzy controller system for quadrocopter was developed on the basis of computer simulations. The mathematical model of quadrocopter and its cascaded fuzzy controller were simulated using Matlab Simulink software. The proposed controller was tested in most frequent flight circumstances: in hover, in rectilinear flight with constant speed, in climbing and in rotation. In all these situations the proposed controller was able to provide foreseeable behavior of the quadrocopter.