Continuous Hidden Markov Model-Based Dynamic Texture Analysis

Texture is the visual characteristic of the object.Dynamic texture is a texture pattern with a certain motion in space-time domain.With the development of artificial intelligence and machine vision,dynamic texture analysis has become a hot topic in basic research.Among them,dynamic texture classification and segmentation as the main means of dynamic texture analysis,play an important application role in medical testing,industrial production,military detection,intelligent transportation and many other areas.The Continuous Hidden Markov Model(CHMM)is composed of two parts that the Markov chain describing the hidden state and the stochastic process describing the probability distribution of the observed values.In the model description of dynamic texture,the hidden state of the Continuous Hidden Markov Model is used to describe the "motion attribute" of the dynamic texture,and the observation vector satisfying the mixed Gaussian distribution is used to describe the "appearance attribute".Accordingly,it is reasonable to describe dynamic texture using the CHMM model and this paper mainly studies the dynamic texture classification method and segmentation method based on Continuous Hidden Markov Model.The main work in this paper is described as follows:1.A dynamic texture classification scheme based on CHMM is proposed.This paper use the gray value sequence of dynamic texture as the observation vector by extracting 16 neighborhoods and 24 neighborhoods gray value,to establish the dynamic texture CHMM model.The modeling principle in this paper is that,the gray scale intensity of the dynamic texture at different moments is regarded as the output of the mixed Gaussian distribution of the model,the magnitude of the gray scale intensity changed with time is regarded as the result of the hidden state transition.On the basis of model description,the dynamic texture classification is carried out by using the maximum likelihood criterion.The classification performance is analyzed by comparing with model description of Discrete Hidden Markov Model.2.A dynamic texture segmentation scheme based on CHMM spectral clustering method is proposed.Considering the numerical underflow problem,the KL-Divergence calculation method of CHMM model is improved and used in the CHMM spectrum clustering segmentation scheme.Then,partitioning the dynamic texture in space and the CHMM model is established for each subblock.The spectral clustering of CHMM is performed by measuring the three distances of improved KL-Divergence,MFS,BP method,to obtain theinitial segmentation result of dynamic texture.Simultaneously,the representative CHMM model of the connected region is obtained by using the k-means clustering algorithm.At last,calculating the probability of the pixel point gray scale sequence that is generated by the representative model,the pixel level segmentation of the dynamic texture is carried out by using the criterion of the likelihood criterion.The segmentation performance of the proposed scheme is analyzed by using the segmentation accuracy index.