Research On Image Denoising Based On Edge Enhancement And Sparse Representation

Image denoising is an important part in the field of computer vision. The goal is to recover the latent clean image from a noisy observation. Recently, image denoising approaches can be categorized as spatial domain, transform domain,and sparse representation based dictionary learning. With the rapid development of the compressive sensing technology, the denoising algorithms based on sparse representation show a good adaptability. By learning or designing an appropriate dictionary to adaptively represent the given patch, the structural characteristics of the image can be better expressed. In this paper, we study image denoising method based on sparse representation theory.Sparse representation and nonlocal self-similarity play a very important role in image denoising. However, it is possible to produce the over smooth or pseudo texture phenomenon for the details of the structure of the image (for example,edge structure) in the application of sparse representation and nonlocal self-similarity. To improve the performance of image denoising, in this paper we propose an edge enhanced and nonlocal sparse representation (ENSR) model which combines Sobel edge detection results, local sparsity and nonlocal self-similarity. The image denoising method can better preserve the edge structure. We use the iterative shrinkage algorithm to solve the l1-regularized ENSR minimization problem. Experimental results show that the performance of image denoising has improved when the edge regularization term is added. ENSR is a good way to protect the edge structure and improves the quality of reconstructed image.In the practical application, the distribution of noise is more complex. We improve the ENSR model to solve the problem of mixed noise. In the sparse coding process, we use the weighted encoding to deal with mixed noise whose distribution has a heavy tail. We propose an edge enhanced and nonlocal sparse representation model for mixed noise (M-ENSR). We use iteratively reweighted scheme to solve the model. Experimental results show that our method has a large advantage in preserving the edge and detail information by comparing with other conventional mixed noises removal methods. And our method has important practical application value.