Research Of Image Denoising Methods Based On Structure Similarity

Recently, more and more researchers dedicated themselves to researching thefields of image processing. These researchers promote the development of imageprocessing technology. In this paper, we study the structural similarity and put forwardsome new models. At the same time, efficient algorithms are proposed for the models.The main content is summed up as follows.1. In order to overcome the weakness that the fidelity term of the classical ROFmodel does not consider image structure and lead to poor visual of image restoration,we propose a new image denoising model (model1) by introducing structural similarityas the fidelity term instead of the original. For the purpose of preserving sharp edges,we put forward another model(model2) on the basis of model1. In model2, anonconvex regularization term is used instead of the classical TV regularization term.Experimental results show that the performance of the proposed denosingmethods(model1and model2) are better than ROF model in terms of visual evaluation.And Model2has an advantage of model1over preserving edge.2. The edge regions of image contain complex structure information and we canhardly find similarity patches.Therefore, the denoising ability of NL is poor in the edgeregions. Then we introduce structure similarity into weight function to deal with theedge regions. Due to structure similarity fully depends on the structure information ofthe image, it is able to distinguish between similar patch and unsimilar patch accuratelyand improve the performance of the algorithm. Experimental results show that thestructure information is preserved better when removing noise efficiently. Therefore, theproposed method can achieve better perceptual quality than NL.3. An improved sparse regularzation method is proposed based on the structuresimilarity, sparse presentation and dictionary learning. In K-SVD, we use structuresimilarity as a measure instead ofL2norm. At the same time, we also present a methodto solve the model proposed. The new model contain both the advantage of sparserepresentation based on dictionary learning and the advantage of that the structuresimilarity fully depends on the correlation between pixels. Experimental results showthat the new method can preserve structures better and therefore achieves betterperformance in visual evaluation.