| Dynamic textures are sequences of images of moving scenes that exhibit certain stationarity properties in time.Dynamic texture is an important visual cue for image analysis.Therefore,it has been applied to some fields,such as facial expression recognition,intelligent transportation system,public security and so on.In the dynamic texture analysis with model based method,the Hidden Markov model(HMM)based dynamic texture characterization has shown a good effect.But,HMM modeling dynamic textures does not take the dependence between spatial pixels into account.However,the texture is a region property.In view of the shortcomings of HMM,the following work is carried out in this paper.1.The dynamic texture classification method using the multivariate Hidden Markov model(MHMM)is proposed.The traditional univariate HMM can not effectively describe the relationship between different pixels of the dynamic texture.We propose the multivariate Hidden Markov model combining the basic theory of traditional HMM with multivariate Markov chain,and complete theory derivations of evaluation,decoding and training.MHMM is a statistical model which can describe changes in time and depict the dependency between adjacent pixels in dynamic texture space effectively.Therefore,it improves the performance of dynamic texture classification.The simulation experiments verify the validity of the dynamic texture classification method using the multivariate Hidden Markov model.2.A dynamic texture segmentation algorithm based on MHMM is proposed.We partition dynamic texture which needs to be segmented into blocks and describe them with MHMM.Some representative models will be selected with the k-mean algorithm.At the same time,we choose Euclidean distance as the similarity measure.The representative models and the dynamic texture to be segmented are used to generate the feature vectors by the evaluation problem of MHMM.Then,the segmentation results can be obtained by clustering the feature vectors with spectral clustering.Finally,the performance of the method is verified by comparison experiment with the segmentation precision as the evaluation standard. |
