Research On Image Denoising Algorithm Based On Sparse Representation

With the development of modern science and technology, image has become the important carrier of information representation and dissemination. The amount of digital image data increases sharply, which is changing people’s lifestyles and poses a major challenge to digital image processing technique in the meanwhile. During the collection, acquisition, and transmission of the images, noise in images introduced by imaging devices, environmental conditions or human factors is unavoidable and has serious effects on image subsequent processing. Image denoising adopts proper noise elimination techniques to remove the noise in an image in terms of image’s noise model. At present, image denoising technology is applied in a large scale in science domain and plays a significant role in many fields such as security monitoring, medical diagnosis, militaryapplications and remote sensing. In consequence research on image denoising technology has an important and profound realistic significance.This thesis studies the image denoising technology based on sparse representation and mainly focus on the dictionary learning algorithm, the image denoising algorithm based on sparse representation, the image denoising algorithm based on nonlocal self-similarity of images and sparse representation model. The contents of the thesis can be divided into three sections that are detailed as follows.Firstly, sparse signal expansion theories mainly include two parts signal sparse decomposition algorithm and dictionary learning algorithm. Dictionary design is a fundamental ingredient in the Sparse-Land model. The performance and computational complexity of the common dictionary learning algorithms still need improvement. After analysising on the merit and demerit of these dictionary learning algorithms, we propose a new method that combines Method of Optimal Directions (MOD) with Approximate K-SVD (AK-SVD) for dictionary learning. Experimental simulations demonstrate the efficiency of our proposed algorithm and its promising performance on the recovery of a known dictionary and dictionary learning for natural image patches.Secondly, the regular image denoising algorithm based on sparse representation exploits the dictionary learning method that trains an adaptive dictionary and calculates the sparse representations of all patches of image to reconstruct the desired clean image. There is an obvious disadvantage that data sample in the regular image denoising scheme is only extracted from the corrupted image to trains an adaptive dictionary ignoring the reconstructed clean image each iteration. To solve this problem we propose a new scheme that involves the optimization:combining the given noisy image and its denoised version to generate an image database devoted to dictionary learning which can be solved by applying block-coordinate minimization algorithm. Experimental results demonstrate its competitive performance in image denoising, as compared with the regular denoising scheme.Finally, we focus on the image denoising algorithm based on nonlocal self-similarity of image and sparse representation model and analyze the nonlocally centralized sparse representation (NCSR) algorithm and patially adaptive iterative singular-value thresholding (SAIST) algorithm in detail. As the common way to measure the similarity between two image patches fails to consider fully their structural similarity, we introduce the structural similarity (SSIM) index into the NCSR and SAIST algorithm to measure the structural similarity between two image patches. The experimental results show that the SSIM can find image patches with the structural similarity and is able to retain the image structure information and can improve the denoising performance compared to the state-of-art denosing algorithms.