Images are inevitably contaminated by noise from equipment and surrounding environment in the process of acquisition and transmission, which will seriously affect the visual effects and the results of the subsequent processing. As the important step of image processing, denoising results will largely determine the effect of image feature extraction, segmentation, compression etc. Therefore, denoising has been the very important research topic in recent decades.Image denoising algorithm can be divided into several stages:The first stage is algorithms processed in spatial domain, including mean denoising, median denoising; the second stage is algorithms processed in transform domain, wavelet transform, Contourlet transform included; the third stage is non-local denoising algorithm. Of course, the algorithms via sparse representation, being recent hot issue, canâ€™t be ignored. Sparseland model achieved significant results in the field of image denoising, super-resolution, face recognition, compressed sensing etc., while there are problems either. This dissertation analyzed the sparseland model, the self-similarity of the images and sparseland denoising model and detail contents include the following aspects:First, basic theory and development process of sparseland model were introduced. The signal representation using sparseland model were analyzed. With the existing algorithms of solving sparseland model, sparse coding and dictionary learning methods were summarized in detail. Meanwhile, the advantages and disadvantages were introduced.Second, the source, measurement of noise and the development process of denoising methods were briefly introduced. The typical denoising algorithm and their existing problems were discussed. Considering the problems of non-local denoising method and traditional K-SVD denoising method, a new non-local denoising method based on sparse representation was proposed, which used the advantage of non-local means denoising method to find self-similarity and denoising effect of K-SVD. Firstly, image blocks were clustered. On one hand, it can make full use of parallel computing to speed up the dictionary learning procedure, which improved the K-SVD denoising algorithm; on the other hand, the separate dictionary learning for the clustered blocks can make the trained dictionaries more targeted, which can maintain the structure information and get better reconstruction effect. Secondly, sparse K-SVD was used to train separate dictionaries and reconstruction. Compared with the traditional K-SVD algorithm as well as non-local means denoising algorithm, the algorithm denoising effect, the structure information remain and the computation were good.Third, a new denoising method based on non-local regularized sparse representation were proposed, which added sparse coding noise regularization in the above method. Firstly, image blocks that similar in structure were clustered by utilizing the idea of non-local denoising. This reinforced the adaptive ability of dictionary because each image block runs dictionary learning independently. Then, structured dictionaries within classes were learned through substituting the K-SVD by sparse K-SVD. Finally, in order to improve the effect of image reconstruction, the sparse coefficient error regularization was brought in to revise the sparse coefficient. Compared with traditional K-SVD denoising algorithm, experiments showed that the proposed method can protect the information of image structure effectively and promote the result of denoising greatly. Simultaneously, without decreasing the structural similarity image measurement (SSIM) value, the peak signal to noise ratio (PSNR) value was very close to the advanced denoising algorithm and sometimes even better than that. |