Методи фільтрації текстурованих зображень у задачах розпізнавання та класифікації

Abstract

The thesis is devoted to solving the problem of textured images filtration for the texture type classification and heterogeneities identification. It was showed, that harmonic series on the basis of eigen fluctuations of texture signal provides an optimal number of parameters for the texture presentation, which is analogues to the eigen vectors decomposition. Texture filtration by autoregression models was improved. For texture periodicity reflection the new condition of two-dimensional parameter vector symmetrization was offered. It improves the filtration of main harmonic components and enforces the sensitivity to factors that destroy texture periodicity. Also, for reflection of constant signal values changes by the model the nonlinear component as image signal squares was introduced. Sensibility of the autoregression model was increased by the additional auto influence signal which was found by variation of the model error. The method of textured background filtration on the basis of the inverse resonance filter model was proposed. The model allows to present the filtration as the inverse problem for definition of the known actuating signal with simple form and to solve it using resonance harmonic vectors decomposition. The method of resonance harmonic estimation in the condition of signal with phase and derivative faults was improved using conditions of linear and unitary symmetry of image model. The problems of filtration optimization in spectral domain were solved by the phase spectrum correction and creation of the optimal, adapted functions basis with eigen kernel of harmonic functions and by using of fast spectral transform, which decreases the operations number in 4 and more times. The method of quasi-periodical textures filtration by principal harmonic components approximation was offered. Method allows to leave weak informative signals undistorted and to recreate them with much higher visualization quality and resolution ability. The software for filtration of the texture as marine surface and marine mammals recognition, biological textures and granulated materials classification was designed.