The Research Of Image Denoising Based On PDTDFB And HMM

Image denoising is a classical subject in image processing. In recent years, image denoising algorithm based multi-scale geometric analysis becomes more and more popular in this field. Multi-scale geometric analysis combined with HMM is a research hot spot at present. Comparison with the traditional denoising method, the HMM based on multi-scale transform domain can fully capture and describe the statistical correlation of the multi-scale transform coefficients in inter-scale and inter-direction, and effectively solve the blur problem of image edge. Therefore it has promising applications to image de-noising.In this paper, after studying the elementary concept of PDTDFB transform and HMM, the dependencies between PDTDFB coefficients and the statistical properties of HMM are paid more attention. Finally two new denoising algorithms based on PDTDFB and Hidden Markov Tree (HMT) are proposed. The main contributions of this paper are summarized as follows.①Performing PDTDFB transform in noise image, then use HMT model to characterize the dependencies of PDTDFB coefficients, thus a new HMT based on the real part and imaginary part of PDTDFB coefficient is modeled. The 2-state Gaussian mixture model is used to approximate the image PDTDFB coefficients margin distribution and the matrix of state transform can capture the dependencies of the PDTDFB coefficients across the scales, EM method is used to train the HMT to get the parameters of HMT of the noise image, and then through subtracting the noise value to get the HMT parameters of original image. Bayesian is used to estimate the PDTDFB coefficients of original image, finally reconstruct the denoised image.②The size of module is not sensitive to small shift in image, which can lead to more accurate estimation in given location and scale, thus the PDTDFB-HMT model based on the model of coefficient is proposed. After performing PDTDFB transform according to the decompose level and direction in the noise image, and calculate module of PDTDFB coefficient and argument,then model the HMT to characterize the module of PDTDFB coefficients. The middle process is similar to dual-tree PDTDFB-HMT, at last recover PDTDFB coefficient according to the relationship between module and argument, reconstruct of signal with PDTDFB inverse transform to get final denoised image. With regard to Gaussian white noise, a theoretical analysis and simulation results show, compared with several other typical denoising algorithms, whether the objective evaluation or subjective evaluation of image denoising, the two new methods raise PSNR at different degrees and effectively remove Gaussian white noise. They can remain edge information and details of original image, especially the method of PDTDFB-HMT based on module. However, the algorithms should be improved to reduce the time cost.