Anchored Fusion Of Monte Carlo Non-Local Means And Low-rank Tensor

The Non-Local Means(NLM)has attracted much research effort due to it superior performances,the basic principle of the NLM is to estimate the current gray value by using the weighted average of all the pixels,which requires too much numerical calculation.Monte Carlo Non-Local Means(MCNLM)through random sampling of similar blocks to build small subset,applying similar patches of small subset to speed up the classical NLM calculation,according to the set of sampling rate can be greatly reduced computational complexity.Due to imperfect algorithm,the use of neighborhood pixels to replace the current noisy gray value with a weighted average easy to blur the image details,local texture is not sharp,at the same time,the residual image contains the lost texture details,by using residual image can effectively make up for the lost structure,remedy small high frequency information.Image patch in the embedded space presents a highly constitutive property,the traditional image processing methods usually drag image patch into a vector.Data processing based on the perspective of the tensor can maintain the structure information of the data,based on the tensor of block structure can describe internal geometry and protect the edges and details.Meanwhile,tensors with similar patches clustering also have low rank characteristics,which is conducive to sparse representation of images.No image denoising method is the best on all image restoration,the classical image restoration algorithm is complementary in both formula optimization and denoising performance.In this paper,we anchored fusion the Monte Carlo non-local means and the low-rank tensor with a frame,then combine the residual image to iterate filter to improve the denoising effect by constantly updating the parameters and image patchs.In this paper,an algorithm based on low rank tensor and accelerating non-local means is proposed for image denoising.The main research work of the paper is summarized as follows:Firstly,we summary the existing image denoising algorithms and image processing method based on tensor systematically,the basic operation of the tensor and tuckerdecomposition are presented in detail,Monte Carlo random sampling similar patches to speed up the nonlocal means algorithm is further study.Secondly,tensor structured form a patch cluster of a nature image,the data processing method of the visual perception based on tensor can maintain data internal manifold structure and low-rank properties,take advantage of tensor structure combined with the residual image information to compensate the missing edge information effectively.Third,to solve the rank of tensor is difficult,we have no exact formula,we have to tune this tensor rank from a number of combinations to realize tensor has a low rank indirectly,not only the core tensor elements are contracted by the hard threshold operator,but also the base matrixs are optimized,finally,the core tensor and the base matrixs are pruned to make the tensor more sparse.Experimental results demonstrate that the proposed model outperforms NLM,K-SVD and BM3 D,in terms of the objective evaluation criteria and subjective visual effects.