Dynamic Texture Segmentation Based On Scale Inter-context Model

Dynamic texture is a series of image sequence repeating in space, changing over time,and having a certain self-similarity in spatio-temporal domain. Dynamic texture analysis shows potential applications in many fields, and as one of the key issues in dynamic texture research, dynamic texture segmentation is getting more and more attention, the study of dynamic texture has become such a hot issue.Dynamic texture segmentation is a segmentation method, which segments natural textures into a number of non-overlapping regions, and texture show different characteristics in different regions, whereas uniformity and consistency is shown in the same regions.Inter-scale context can make full use of the relationship between the grade of different scales,to depict the "movement" and "appearance" features shown in dynamic textures. Therefore,wavelet domain inter-scale context model based dynamic texture segmentation is proposed in this paper. The main contents of this paper can be concluded as follows:1. We propose wavelet domain Markov chain contextual based dynamic texture segmentation algorithm. Through wavelet transform, the strong dependence relationship between each sub-band in the same frame of image or between sub-band of adjacent scale,was carried out, improving the characterization ability of dynamic texture. The label field model using the inter-scale Markov random field model with inter-scale non-causal Markov random field model, and the feature field model using Gauss-Markov field model is presented,the relationship between each wavelet coefficient vector and its adjacent wavelet coefficient vector in the same scales was considered in the neighborhood interaction parameter matrix.The results show that the proposed algorithm has nice segmentation performance of dynamic texture.2. We propose a novel dynamic texture segmentation algorithm based on Markov random field energy contextual model. According to the spatio-temporal neighborhood systems and multi-scale random field model, the Label field of neighborhood systems and energy function is determined. In this paper, the observation field was described by gauss distribution, forming Markov random field dynamic texture segmentation by the method of multi-scale random field model, to segment dynamic texture by using the rule of maximum a posterior. Finally, comparing with the simulation results of the proposed approach in existing literature, the experimental simulation results the method to obtain better segmentation results.