Research On Image Denoising Approach Based On Dependencies Of Wavelet Coefficients

As an important medium of the human visual information transmission, image is often influenced or interrupted by a variety of noise in the course of transmission. These degraded images will have a negative impact on the following-up image processing (such as segmentation, compression, feature extraction, and pattern recognition, etc.). Therefore, the image denoising becomes a very important job for the image-preprocessing. The image denoising is to remove the noise from the noisy image, at the same time to preserve image detail and feature. The key of the research of image denoising is to improve the signal to the noise ratio, and to highlight the expectations of the image features. However, it is difficult for traditional image denoising methods to reach a ideal trade-off between noise suppression and detail preserving and to obtain good results.With the rapid development of wavelet analysis theory, people pay their attention from spatial domain to wavelet domain. Due to the characteristics of the wavelet such as: multi-scale, multi-resolution analysis, and the improved denoising effect, the image denoising method based on wavelet becomes an important research topic in the field. The main tasks in this paper include the following three respects:(1). Wavelet transform in image denoising occupies a very important position in the field. Fist of all, a series of characteristics of wavelet transform was enumerated in this paper. The development of wavelet denoising history and current situation were introduced in detail. The noise model, the noise variance estimation and the evaluation criteria of the image denoising performance were introduced. The paper focused on threshold selection in image denoising. Several classical thresholds and the corresponding algorithms were analyzed. Some important conclusions were obtained from the simulation experiments. Orthogonal wavelet transform lacks the translation invariance,so the classical thresholding denoising algorithm could be improved through importing the shift-invariant method. This modified algorithm could reduce the pseudo-Gibbs effects of the denoised image edges, and optimize the denoising performance.(2). Based on the studies of intrascale dependencies of wavelet coefficients , the paper proposed a new image denoising algorithm using normal inverse Gaussian distribution model for the coefficient and the context model for coefficients categories. Classical statistical model of wavelet coefficients was systematically studied. The paper focused on the normal inverse Gaussian model. The parameters formula of Bayes maximum posteriori estimate were derived. The classification of wavelet coefficients based on context model was discussed in detail. The paper researched on the moment estimation for parameter estimation. The statistical model introduced in this algorithm could comprehensively describe the intrascale dependencies of wavelet coefficients. The classification method based on context modeling could also embodiment the the intrascale dependencies of wavelet coefficients, with which the shrinkage function became more adaptive and the denoising result was improved.(3). Based on research of the interscale dependencies of wavelet coefficients, the paper proposed a linear minimum mean square error estimation of image denoising in vector space. Decomposition method based on Nonsubsampled and the advantages of redundant wavelet transform was expounded. Context classification model applied to the vector space was deeply studied. The LMMSE algorithm based on Bayes criteria was introduced. The paper focused on a new algorithm based on the LMMSE algorithm applied to the vector space. The new algorithm could resolve the fuzzy edge denoising, and it had greater improvement in the vision and signal to noise ratio than the use of intrascale correlation.