Bogucharskiy S. Methods and models of tolerance clustering in image collection

The thesis is devoted development of models and methods of visual information segmentation in image space on windows processing basis. Matrix modifications of clustering methods, viz CLARANS for video streams processing in very large video data bases and DBSCAN for image segmentation with fuzzy boundaries between arbitrary shape regions, are introduced. Window methods based on data distributions and recurrent optimisation, that allow to form clusters with any shape at high noise level, have been proposed. Matrix modification of X-means and also modifications of J-means and fuzzy J-means, which provide possibilities to ensure reaching of deeper extremum of the accepted goal function of segmentation quality, is offered. The method of semicontrolled matrix vector quantization is developed for textures analysis (segmentations of image sequences), allowing to process video data in sequential mode both learning with the teacher and selflearning. Results of experimental investigations are discussed.