Mixing Models Based Dynamic Texture Classification

Dynamic texture,which is also known as temporal texture,is a kind of visual pattern of image sequences that are spatially repetitive and temporally varied,as well as contain some certain spatiotemporal stationary properties.With the development of science and technology,the efficiency of obtaining dynamic textures has become higher and higher while people have more and more ways to obtain dynamic textures.Therefore,the research of dynamic texture analysis is attracting more and more attention when people are facing massive dynamic texture data.A quite important research content of dynamic texture analysis is the classification of dynamic textures,which has an extremely widespread application prospect in military,medicine,transportation as well as other fields.Wavelet transform is frequently used in texture classification domain.As complex wavelet transform could efficiently compensate the defects of traditional real wavelet transform based method,such as shift sensitivity,poor selectivity of orientations and the lack of phase information,thus the main research in this thesis is the realization of dynamic texture classification in the complex wavelet domain.Popular as well as advanced research methods in the domain of dynamic texture classification are summarized,also the process of dynamic texture classification is detailedly analyzed in the thesis.On the foundation,this thesis proposes a new method for dynamic texture classification by finite mixtures of distribution models in the complex wavelet domain.The main work of this thesis includes the following aspects:(1)Based on characteristics of the relative phase information in the dynamic texture,a new dynamic texture classiciation method is proposed in the thesis.The distribution for phase information of complex wavelet coefficients in detail subbands of the dynamic texture is approximately uniform,which cannot yield any useful information of the dynamic texture.Aiming at relative phase information of complex wavelet coefficients in detail subbands of dynamic texture,the finite mixtures of Von Mises distributions(MoVMD)model and the finite mixtures of Wrapped Cauchy distributions(MoWCD)model are proposed for dynamic texture classification.Corresponding parameter estimation approaches based on the EM algorithm are also presented in the thesis.The obtained model parameters are accumulated into a feature vector to describe characteristics of the relative phase information of the dynamic texture.To measure the similarity between dynamic textures,variational method and matching-based method are introduced into the classification method.Furthermore,the experimental results on two popular benchmark dynamic texture datasets(UCLA and DynTex++)demonstrate the effectiveness of the proposed methods.(2)The finite mixtures of Gumbel distributions(MoGD)model and the corresponding parameter estimation approach based on Expectation-Maximization(EM)algorithm are proposed in the thesis.Each detail subband is divided into nonoverlapping blocks.Then,the median values of complex wavelet coefficient magnitudes of the nonoverlapping blocks are modeled with MoGDs and corresponding parameter estimation approach.The obtained model parameters are accumulated into a feature vector to describe characteristics of the median values of complex wavelet coefficient magnitudes of the dynamic texture.To realize the classification,a variational approximation version and a matching-based approximation version of the the Kullback-Leibler divergence(KLD)are used to measure the similarity between two MoGDs.Extensive experimental evaluations on two representative benchmark dynamic texture datasets(UCLA and DynTex++)demonstrate that the proposed MoGD based classification method shows better performance in comparison to recent state-of-art methods,such as VLBP,DFS and so on.