Texture Image Feature Extraction Based On Multi-Scale Transform Domain Hidden Markov Tree Model

Compare with the approximation properties of wavelet analysis, multi-scale geometric analysis is just as wavelet analysis in relation to Fourier analysis.Curvelet transform is multi-scale, the introduction of the direction parameter make it highly anisotropic, and it describes image edge excellently only with fewer non-zero coefficient. Curvelet is able to remove noise thoroughly and can protect edge well, but the staircase effect is easily brought about. Image de-noising based on partial differential equations (PDE) cuts a great figure and is good at image smoothing in image processing. Smooth regions are protected well and the staircase are not produced by Four-Order partial differential equation, but it has low de-noising rate and is bad at protecting edge. In order to de-noising for noisy image and help improve feature extraction and the efficiency of image retrieval, a new model is proposed in this paper by combing Curvelet and LLT model in Four-Order partial differential equation with weight function. The experiment demonstrates that the model could preserve the advantages of Curvelet and LLT, at the same time, PSNR and visual effect are superior to either of them.Combined by multi-scale and directional filters banks, Contourlet transform has the character of anisotropic, and its decomposition coefficients are non-Gaussian and persistence. At present, in order to take advantage of correlation to descript texture image, Contourlet domain hidden markov tree model concerns that neighbor nodes of the parent node affects child node, weighted Contourlet domain hidden markov tree model is proposed in this paper. This model not only considers parent state, but also the influence of parent's brother is used by weight. The state transition matrix represents their influence. That makes the new model can descript the internal relation between Contourlet coefficients and HMT more accurately. Kullback-Leibler Distance is used to measure the similarity between images, and the experiment results demonstrates that the average retrieval rate using new model is about 7% to 46% higher than that of CHMT method.