| With the development of economy, advancement of society and steady improvement of people's standard of living, texture classification has been applied to many academic subjects and every aspect of people's everyday lives. Due to this era of information explosion, traditional static classification has showed its non-effectiveness, dynamic texture classification method has attracted much of people's attention.This thesis is mainly focused ondynamic texture classification. At first we introduce the research significance and current status of dynamic texture classification, and make a study of basic theory on wavelet analysis, pointing out the rationality of wavelet analysis being applied to dynamic texture classification. Our work in this paper thesis is as follows:(1) Anew wavelet transform method is proposed which conduct 1-D wavelet decomposition along the time axis of video files, and then sub-band can be processed using 2-D spatial wavelet transform, thus giving consideration to dynamic local features. The new classification method is based on the fact that the sub-band coefficients of wavelet transform follow the Generalized Gaussian Distribution (GGD). Compared with traditional methods based on the mean and standand deviation, experiment resulsts using KNN and SVM verify the effectiveness of our new method, we use the method of feature selection in order to avoid the high feature dimension brought about the amount increase of the calculation and the deterioration of classification performance.(2) We make a research on dynamic texture classification based on Dual-Tree Complex Wavelet Transform (DT-CWT). Amplitude of complex coefficients of sub-band decomposed by DT-CWT follow the probability density of gamma and lognormal. The thesis takes parameters following these two kinds of probability density as dynamic texture feature. Throuth experiments implementing 2-level, 3-level DT-CWT and the two classification methods mentioned vadility of the new method can be comfirmed. |
PDF Full Text Request