Dynamic Textures Segmentation Based On Multiresolution Hidden Markov Random Field

Dynamic texture is a spatially repetitive, time-varying visual pattern that forms an imagesequence with some spatio-temporal stationary properties. Dynamic texture is a veryimportant feature in the video, which provides important information for video processing andanalysis. Dynamic texture plays an important role in texture analysis, and which could bewidely applied in many fields, such as military affairs, industries, medical treatments,intelligent transportation, meteorology, public safty and so on.Dynamic texture segmentation is based on a constant space-time statistics, dividing anatural scene image sequence into several regions of the mutually overlaping, each region hasa uniform texture. The spatio-temporal hidden Markov random field model can effectivelydescribe the "movement" and "appearance" feature of dynamic texture, so, this paper aims tostudy the dynamic texture segmentation based on spatio-temporal hidden Markov randomfield model (STHMRF). Then we extend STHMRF to the wavelet domain combined with thewavelet multi-resolution features, and format a multi-resolution space-time hidden Markovrandom field model. The main contribution in this paper is as follows:1. Spatio-temporal hidden Markov random field (STHMRF) based dynamic texturesegmentation. According to space-time characteristics of dynamic texture, we establishneighborhood system and energy function of MLL model in marking field, and describeobservations with Gaussian Markov Random Field. Then we form Spatio-temporal HiddenMarkov Random Field model and use algorithm of EM and MAP to estimation parameter andsegment dynamic texture at the same time. The simulation results show the proposedalgorithm has obvious advantages.2. Scalar multi-resolution space-time hidden Markov random field (scalar MSTHMRF)based dynamic texture segmentation. According to the advantages of the wavelet coefficientsfitting to the Gaussian distribution, we put forward the scalar MSTHMRF dynamic texturesegmentation algorithm. The method uses the implicit relationship of wavelet coefficients indifferent scales of the same direction, and initializes the label field of the low-level detailsubbands by the STHMRF segmentation result of senior detail subbands. Then, we initializeoriginal image label field by integrating segmentation results of low-level details sub-bandand use STHMRF segmentation to get the final results. 3. Vector multi-resolution hidden Markov random field (vector MSTHMRF) baseddynamic texture segmentation. On the basis of the scalar MSTHMRF segmentation algorithm,the vector MSTHMRF algorithm considers the wavelet coefficients relationship of same scalesub-band. This method look vector composition of the each subband coefficient in the sameposition of the same scale as a seven-dimensional observation field, and then use the vectorMarkov random field model to segmentat step by step.