Image Denoising Based On Sparse Decomposition

Image denoising is one of the key problems in image processing and the basis of image subsequent processing. Many image denoising methods have been developed based on the feature of noise, which have good performance in different conditions. However, these methods are always based on the statistic features of image signal and noise. Unfortunately, these features can not be known in advance in real applications.According to the different characters of image and noise in sparse decomposition, an adaptive image denoising method is proposed in this thesis. The research results are as following:As one traditional method for denoising, mean filtering has some short-comings. To overcome these drawbacks, one method is proposed to create different types of templates. For each noise level, only one kind of template is fit to obtain the optimal denoising result. The effect of this method is compared with that of the following method based on sparse decomposition.The anisotropic atom dictionary that is more suitable for image sparse representation is chosen according to the feature of the matching pursuit image sparse decomposition and human vision system. Based on study of the different behavior of image and noise in sparse decomposition, the difference between image and noise is identified. According to the different coherent ratio between image, noise and over-complete dictionary, image information and noise are distinguished. One image adaptive filtering is realized by taking coherent ratio threshold as the condition of ending the sparse decomposition process. Experimental results prove that the select of coherent ratio threshold is independent of image types or noise level. So the way of denoising presented here is adaptive.The effectness of image denoising between sparse decomposition and best smooth template are compared. According to the vision effect, image denoising results based on sparse decomposition is better than that of best smooth template.