Image Denoising Based On Regional Division And Dictionary Learning

The image denoising technical is an important research topic in the field of digitalimage processing. As a hot research topic, the sparse representation theory has broughtnew ideas for image denoising methods in recent years. The domain of sparserepresentation from orthogonal basis atoms to the redundant dictionary, to trainingredundant dictionary by learning adaptively, reflecting the sparse representation theorytends to redundancy and adaptability gradually. The denoising effect of the method oftraining redundant dictionary for image denoising through sparse representation is good,but such algorithms also have limitations, such as the denoising algorithm based onK-SVD appears fuzzy in the place of weak textures and weak edges, artificial traces willalso be found in the smooth regions, at the same time, when there has much noise in theimage, the denoising effect is not so ideal.In order to solve the problems above, we propose a method of image denoisingbased on regional division and dictionary learning after making research on the BM3Dalgorithm and the non-local means algorithm. The innovations of this work are asfollows:1. We proposed a method of image denoising based on the combination of K-SVDand BM3D. When the image denoising algorithm based on K-SVD solves the sparsecoefficients by the OMP algorithm, the standard deviation of the image blocks in weaktextures and weak edges is smaller than the threshold given by the OMP algorithm, theycan not use the dictionary atoms to represent, but just use the local mean method toachieve denoising, so there will be the fuzzy phenomenon. We use the BM3D algorithmwhich is better for texture and structure signals to denoise the image blocks whosestandard deviation is less than the threshold given by the OMP algorithm. It solves thefuzzy problem in the place of weak textures and weak edges caused by the denoisingalgorithm based on K-SVD efficiently.2．We proposed a method of image denoising based on regional division anddictionary learning. Firstly we use the Primal sketch model and the statisticalcharacteristics of the image blocks to divide the image into structural regions, textureregions and smooth regions. We use the denoising method based on the combination ofK-SVD and BM3D to denoise the structure regions and the texture regions. As theRidgelet redundant dictionary is suited for the signals in the area of edge and contourand the DCT redundant dictionary is suited for the signal in the area of texture, we choose them as the initial dictionary, so as to represent edge and texture information ofthe image more effectively and speed up the convergence rate of the algorithm. And thesame time, we use the method of Non-local means to denoise the smooth regions, itsolves the problem of artificial traces in smooth regions caused by the image denoisingmethod based on K-SVD. The result of the simulation experiments shows that thismethod improves the smooth area denoising effect and keeps the image textures andedges effectively compared with the image denoising method based on K-SVD.