Research On Sparse-based Image Denoising Algorithm Based On Regional Division And Sparse Representation

Image denosing has been an important research aspect in digital image processing.In recent years,the sparse representation is one of the research focuses.By selecting or designing proper extra dictionary,sparse representation can effectively capture major features in the image and describe it in a flexible way as far as possible.In the method of traditional image denoising,image should be projected to some certain transform domain,where signal and noise can be seperated apart.However,signal and noise can be separate in such a spare domain,which will cause a loss of the original image information in denoising.On the contraty,using the method based on the theroy of transform domain,signal and noise can be completely separate from each other because of noise is not the component of spare in the signal.The main works are as follows:(1)Learning dictionary denoising algorithm based on K-SVD.In this paper,the basic theory of traditional denoising methods are analyzed,such as Gaussian filter,Bilateral filter,Wiener filter,Non-local Means(NLM)denoising method,and so on.The sparse representation theory and the basic model of sparse denoising are studied,and the two links of sparse decomposition and dictionary learning are analyzed.In the image denoising,compared with the traditional dictionary,K-Singular Value Decomposition(K-SVD)learning the dictionary is obtained through machine learning,including the image of its own characteristics,it has better denoising effect.Experimental results show that the adaptive dictionary denoising effect is generally better than the Discrete Cosine Transform(DCT)and Global Dictionary(GD).(2)Denoising based on regional division and sparse representation algorithm is proposedFor a natural image contains both non-smooth region and smooth region,this paper reprensents the algorithm of denoising based on regional division and sparse representation.First of all,through the study,that different sparse representation denoising algorithm based on the same image of different types of regions of the denoising effect is different,that different algorithms for different regions have different denoising effect.Then using image edge information measure of directional information measure,the two parts of an image classified as non-smooth regions and smooth regions,represents a combination of different sparse denoising algorithm,to restore the original image.Experimental results show that the proposed algorithm is better than using the same sparse representation denoising algorithm.(3)An improved based on feature CSR algorithm is proposedThis paper presents an improved clustering algorithm based on sparse representation denoising.In conventional clustering based on sparse representation denoising(CSR)dictionary acquisition process,the raw data is sampled under way in multiple scales,training related to the image dictionary.But the image of the next sample will be missing part of the image information,and in the subsequent image block clustering process will cause a block of information does not match,the image denoising will eventually have a bad effect on the results.By using non-sampling Difference of Gaussian pyramid algorithm,without changing the size of the image size to achieve multi-scale representation of the image,improve the accuracy of image block clustering,thus improving the accuracy of principal component analysis,so that the image better able to express the image dictionary feature.Experimental results show that the improved CSR algorithm has better denoising effect than K-SVD local denoising and original CSR algorithm.In this paper,the PSNR value of the restored image is improved by 0.8dB.