Multi-band Image Denoising Based On Nonlocal Self-similarity Modeling

In the process of image acquisition,due to the thermal effect of the sensor and other reasons,the final image will be polluted by a certain degree of noise,which includes Poisson noise,strip noise and so on.Therefore,image denoising,as an image processing task in the low level of computer vision,has always been one of the research focuses in computer vision.On the one hand,it plays an indispensable role in the underlying processing tasks and has a great relevance with other low-level tasks(such as deblurring).On the other hand,as a data preprocessing step of the high and middle level image processing tasks(such as detection and segmentation),it guarantees and improves the performance of related tasks.Most of the image denoising research focused on Gaussian additive white noise,while there are also some special work for salt and pepper noise.One of the main challenges of image denoising is that many important information is lost in the process of image degradation,which makes image denoising become a highly ill-posed inverse problem.Reviewing the classical image denoising algorithm,according to the processing object,it can be simply divided into spatial domain and frequency domain algorithm.The denoising algorithm in frequency domain is mainly to transform the image to frequency domain for a series of processing such as threshold or wavelet base;while the algorithm in spatial domain is to use filter on the image matrix directly;in addition,some algorithms that combine the two domain operations,such as BM3 D,have been developed and have been the benchmark of denoising algorithm.In addition,due to the variety of image types,multispectral and hyperspectral image denoising also has a wide range of research as well as gray and color image denoising,these images are often composed by dozens or hundreds of channels(bands),and contain rich spectral information,which has a great application in the spatial geography such as terrain classification.If the gray image denoising algorithm is directly applied to each channel(band)of color or multispectral image,the correlation between channels or spectra is often ignored,which makes this part of prior information wasted and affects the final denoising performance.Based on this observation,in addition to the color space conversion to eliminate the correlation,some denoising algorithms,such as CBM3 D and BM4 D,have been developed for color multichannel image.Low rank matrix approximation(LRMA)has wide range of applications in computer vision and drawn much attention in recent years.The rank minimization is NP hard and is tough to resolve directly.Therefore,it is necessary to add some regular terms to constrain the solution space according to the prior information,and the problem become easy to solve.So how to design the prior information regular terms has become a key step affecting the performance of the algorithm.The typical nuclear norm minimization(NNM)is often given to solve LRMA,but it's apt to over-shrink the rank components due to the same threshold.To address this problem,we propose a flexible and precise model named multi-band weighted lp norm minimization(MBWPNM).We reformulate it into the non-convex lp norm subproblems under certain weight condition and solve these subproblems via a generalized soft-thresholding algorithm.We then adopt MBWPNM to image(gray,color and multispectral)denoising.The proposed MBWPNM not only guarantee more accurate approximation with a Schatten p-norm in case of change of noise levels,but also considers the prior knowledge where different rank components have different importance.Extensive experiments on additive white Gaussian noise removal and realistic noise removal demonstrate that the proposed MBWPNM achieves a better performance than several state-of-the-art algorithms.