Wavelet Based Dynamic Texture Segmentation Using HMT Models

Texture analysis is a basic research subject in the field of image processing. Dynamic texture is a category of texture, and acts as an important part of texture analysis, has been a hot research topic in the field of computer vision and digital image processing in recent years,and also has a broad application on military, industrial, medical, intelligent transportation,meteorological remote sensing, public security and other fields.Multi-Scale Analysis is also known as Multi-Resolution Analysis, is arising as a new signal analysis technique recently. As a signal analysis tool, Multi-Scale Analysis provides a new perspective of research subject, which describes the scene from the perspective of different scales, analyzes and solves problems by using different scale, analyzes effectively by extracting the image or video features. Wavelet transform is applied as Multi-Scale Analysis basis, and highlights some of the features in the signal adequately for its nice time-frequency properties. Markov theory as a powerful tool to describe local area distribution, combines with properties of wavelet domain spatio-temporal transformation effectively, promote the development of wavelet-domain model, and play main action to accurately depict non-stationarity characteristic of the signals. Wavelet-domain hidden Markov tree (HMT)model can take full advantages of the dependences intra-scale and inter-scale of wavelet coefficients, which showed remarkable performance in the field of texture analysis, etc.In the light of above theoretical basis, this paper mainly studies wavelet based dynamic texture segmentation method using HMT model, the main work as follows:1. Proposing 3-D wavelet-domain LMM-HMT model. We extend the HMT to the region of three dimensional spatio-temporal region, modify the EM algorithm for parameter estimation and improve the segmentation process to meet the requirement of three dimensional dynamic texture processing. For the training and raw segmentation step, consider the distribution law of wavelet coefficient of texture image-the peakier in the center and heavier tail of Laplacian distributions, matching Laplace mixture distribution of HMT model,corresponding formulas for estimating parameters were derived, combined with the distribution of wavelet coefficients in texture image. For the multiscale fusion step,performing an interscale fusion method based on 26 neighborhoods context vector. The 3 frame image textures based 26 neighborhoods context vector model was generated for interdependencies between the frames. The finally effects of image segmentation have been improved in using of Laplacian mixture distribution and parameter estimation algorithm in the pixel-level segmentation step.2. Proposing 3-D wavelet-domain HMT-3S model. We establish the HMT-3S model on 3-D wavelet-domain by tying the wavelet coefficients of seven sub-bands, put forward the model parameter training method, proceed to likelihood function calculation, and multiscale fusion method for HMT-3S. Implying the improved algorithm for 3-D wavelet-domain HMT model to HMT-3S model, the segmentation effect gets better and has a nice performance.