Zinko T. Analysis and synthesis algorithms for recognition and classification and their application in processing speech signals and images

The work is dedicated to the development of mathematical tools of analysis and synthesis of classification and recognition of speech signals and images systems. The idea of feature vectors design based on formants components of spectrograms was proposed and implemented on the stage of recognition algorithms constructing for the audio-signal processing. Two stage clustering algorithm with the dichotomy as a first stage and with linear discrimination on the second step was proposed and implemented in the thesis. This algorithm was designed for the fixed collection of the words. Linear discrimination is applied for enriched classes defined by dichotomy. The dichotomy advantage is the high rate of classification whereas linear discrimination provides the stability of algorithm. Special peculiarity of the algorithm proposed is systematic using of the Moore-Penrose pseudoinvers on the base of singular value decomposition. It is fully valid for the algorithm of "algebraic filtration". The development of inhomogeneous "classes accordance distance" is proposed in the thesis for the case of Euclidean matrix spaces. New results related abstract Hough-pair are proved in the thesis and applied to the solution of recognition problem for the plates.