Spatio-Temporal Markov Random Field Based Dynamic Texture Segmentation

With the development of Science Technology and advancement of society, the research on dynamic texture is a hot issue in the texture analysis. Dynamic textures are representations of such textured surfaces with repetitive, time-varying visual patterns which form an image sequence that exhibits certain stationary properties in time.At first, this paper introduces the research significance and current status of dynamic texture, and analyses the reasonableness of Markov field theory analysis applied to dynamic textures segmentation, then, proposes a space-time Markov random field model for dynamic texture segmentation, and this model is extended to the wavelet domain. The main contributions in this paper is as follows:(1) Spatio-temporal Markov Random Field based dynamic texture segmentation. According to space-time characteristics of dynamic texture, we establish neighborhood system and energy function of MLL model in marking field, and describe observations with Gaussian Markov Random Field. Taking account the final segmentation results and computational complexity, we use the maximum a posteriori criterion (MAP) and local optimization strategies for the parameters estimationand the dynamic texture segmentation simultaneously. Finally, simulation results show that the proposed algorithm is reasonable and effective.Compared to the simulation results of the proposed method in [39] and [40], the results show that this approach has obvious advantages.(2)We propose a dynamic texture segmentation algorithm based on space-time Markov Random Field in wavelet domain. According to the characteristics of the multi-resolution wavelet transform and wavelet coefficients obeying Gaussian distribution The algorithm carrys on a space-time wavelet transform for dynamic texture. First,the low frequency sub-band is segmented, we get the preliminary segmentation results as the initial value at the same level detail sub-bands split, and then we make a segmentation by using Markov random field model; the segmentation results is as the initial value at the first level detail sub-band split, and then we make a segmentation using models based on Markov random field again;segmentation results are fused to form the initial value of original image label field, at last, we make a segmentation by re-using model based on Markov random field. Experimental results show that the algorithm is better than segmentation algorithm of space-time domain based on Markov random field.