In this paper, a segmentation method for grayscale images in the framework of the approach based on region merging is proposed. The initial segmentation is obtained by finding edges in the image and its splitting into regions of a given shape. The adopted segmentation criterion is formulated in terms of the least squares fit to the image intensity function and this can be accomplished by finding an optimal partition with respect to the introduced information measure.

The solution of the "structure-property" based on the molecular graphs descriptorsselection with k-NN classifier is proposed. The results of comparing the construction ofpredictive models using the search and without it are given. The high stability of the classifier function construction quality is tested using the real sample of the molecular structures.

A novel approach is proposed to constructing a Bayes classifier in a multidimensional space of fea tures by using treestructured Gaussian mixtures as estimates of classconditional probability density func tions. A training procedure is developed for the classifier that is reduced to finding numbers of mixture com ponents and their thresholds in order to realize rejections for the given classes. The mixture parameters are optimized by a crossvalidation method. Classification error rate is estimated on a set of 3D vectors of textual features of a monochrome image. Comparative error rates are obtained for classifiers that use histograms, individual Gaussian densities, and Gaussian mixtures constructed using the EM (expectationmaximization) algorithm. The practical application of the developed classifier is illustrated by results of image segmentation for a satellite picture. The image represents a fragment of the Earth surface and it is obtained using the Google Earth program.

The paper proposes various approaches to classifying signbased representations of images based on distance functions. Any image is represented as a set of features describing differences in brightness. The construction of a distance function is proposed using classical functionals of information theory: the Shannon entropy and the Kullback–Leibler distance. It is shown that the Bayes classification in the case of independent features can be also described by distance functions. In the last section, the proposed approaches are evaluated using a face detection problem.

In this paper, successful detection of P300 wave embedded into electroencephalogram (EEG) data is aimed. Detection performance of a previously applied method is increased by using proper pre-processing scheme. Development in the detection performance in terms of overall classification accuracy is presented in a detailed manner. The proposed method is highly capable of detecting the minor differences and can be applied to various classification problems.

 

 

Measures and functionals of global asymmetry of noisy and noisefree images are axiomatically introduced. Explicit expressions are obtained that make these functionals applicable for determining the symmetry axes of noisy images. It is shown that some asymmetry functionals are unstable against noise levels of images; i.e., the symmetry axis obtained using these functionals may deviate significantly if the signalto noise ratio is large. Sufficient and necessary conditions are obtained under which the symmetry axes calcu lated using asymmetry functionals remain unchanged.

A new approach to analysis of the molecule–descriptor matrix in the structure–property problem,based on the fuzzy cluster structure of the training sample, is developed. Methods for constructing fast pre diction rejection rules and for the search the outliers in a training sample are described. To that end, a special space ofeasily computed descriptors is introduced. Optimization of the classifying function with respect to the param eters of fuzzy classification is considered. Prognostic models with a high quality of prediction, based on thisapproach, are proposed. Comparison of models is performed, which shows the efficiency of the describedmethods

 A new approach for analyzing the “molecule–descriptor” matrix for the QSAR problem (Quantitative Structure–Activity Relationship) based on a fuzzy cluster structure of the learning sample is presented. The ways for generating fast rules for refusing prediction and searching the spikes in the learning sample are described. For this purpose, a special space of descriptors, simple for calculation, is introduced. The ways for optimizing the discriminant function according to fuzzy clustering parameters are examined. Highly predictive models based on the presented approach have been generated. The models are compared, and the efficiency of the described methods is revealed.

This paper considers an approach to solving the problem of binary classification of objects. This approach is based on representing one of the classes by a sequence of Gaussian mixtures with further introduction of threshold decision rules. A method of constructing hierarchical sequences of Gaussian mixtures using the partial EM algorithm is proposed. We compare classifiers that use single Gaussian mixtures, cascades based on sequences of independent mixtures, cascades based on hierarchical sequences of mixtures, and classifiers that use trees of Gaussian densities for decision making. The theoretical estimates of computational costs for these classifiers are provided. The classifiers are tested on simulated data. The results are presented as the relations between the computational cost of classification and the obtained values of error criteria.

Problems of identification of plane unclosed curves are considered. Methods are proposed that allow one to classify graphic objects invariantly to affine transformations. An answer is given to the question on the types and the quantity of features that are needed to construct a mathematical description of curves for the recognition of an unclosed contour of an object. Metrics are introduced on the basis of which one can identify unclosed curves. The quality of identification on the basis of the metrics introduced is analyzed.

In this work we discuss methods for image ringing detection and suppression that are based on the sparse representations approach and suggest a new ringing suppression method. The ringing detection algorithm is based on construction of the synthetic dictionary that is used to represent ringing effect as a sum of blurred edge and pure ringing component. This decomposition enables us to estimate image ringing level. We analyze two ringing suppression methods. First method is based on learning joint dictionaries and shows good performance for the whole image on average. However for high ringing levels the performance of this method decreases due to the influence of the ringing artefact on the sparse representation parameters. The second method is based on separate learning of natural images dictionary and pure ringing dictionary and it does not suffer from this problem. In this article we present a new ringing suppression method that is based on the method using separate dictionaries. The method works best in the areas of edges and for higher levels of ringing effect.

Proposed and developed a method for solving the “structure property” problem, which is based on an adaptive choice of the description of molecules and the automatic selection of feature space in accordance with the characteristics of the training sample. Solved the problem of combinatorial explosion using Group Method of Data Handling. Used the clustering of objects in the training set to improve the predictive ability of the model.