Non-local Transform Domain Image Denoising Based On Joint Sparse Representation

Image denoising is a widely studied problem in the field of image processing.It is also a necessary process before image segmentation and recognition.How to restore a true original image from a degraded image has always been a hot spot for many experts and scholars.In the field of denoising,wavelet theory is favored by many scholars due to its low entropy,multi-resolution and decorrelation.However,the traditional wavelet threshold denoising algorithm still needs to be improved for separating from noise and preserving edge detail information.In recent years,the theory of sparse representation and non-local self-similarity has been widely used in image denoising,by using the non-local self-similarity of image pixels and designing a reasonable dictionary for sparse representation,which can effectively reduce the sampling cost of image data to obtain the image features,thus enhancing the image denoising performance.Although sparse representation and non-local similarity have achieved good results in the field of image denoising,how to effectively combine the two priori knowledge and improve the denoising performance has been a difficult problem.The block matching 3D collaborative filtering(BM3D)effectively combines the above two theories and has astonishing performance in denoising performance and effect.It is even considered as the best denoising method.Aiming at the limitation of denoising performance of traditional denoising algorithm,this paper studies the sparse representation theory and nonlocal method deeply.By adding the correlation between the real coefficients and sparse coefficients in the sparse representation model,Improved joint sparse representation model,effectively enhance the image sparsity.After analyzing the defects of BM3D algorithm,especially when the noise intensity is large,the performance of the denoising is drastically reduced.In this paper,the wavelet threshold denoising algorithm is fully researched,and the improved scheme of threshold and threshold function is proposed,which is applied to Block matching three-dimensional transform two-dimensional wavelet transform,effectively enhance the denoising performance,making the threshold value can be adjusted according to the noise intensity and image decomposition scale changes.In order to solve the problem of denoising model regularization,this paper first solves the regularization coefficient in the denoising model by Bayesian estimation method,and effectively balances the tradeoff between fidelity and regularity.In this paper,the denoising performance is further optimized and the convergence speed is improved by combining the Split-Bregman iterative algorithm with the denoising model proposed.Finally,compared with other denoising algorithms,the results show that the proposed denoising based on joint sparse representation can not only remove more noise,but also retain the details of edge information and improve effectively Image stability and robustness.