Two Hyperspectral Image Denoising Algorithms Based On Weighted Schatten P-Norm Minimization

Hyperspectral images are widely used in mineral identification,medical analysis,food safety,etc.In the process of acquisition and transmission,the noise will inevitably appear.The appearance of these noises will lead to corresponding deviations in subsequent identification and analysis.In order to improve the accuracy of subsequent applications,denoising has become an extremely important step in the processing of hyperspectral images.One of most advanced methods for hyperspectral images is nonlocal low-rank denoising.For the low-rank approximation problem,the previous work proved that the non-convex low-rank method can get a better low-rank approximation solution.Based on this idea,this thesis presents two new non-convex hyperspectral image denoising algorithms: the WSN-GNL algorithm and GNS-ITboost algorithm.For hyperspectral images with only additive noise,this thesis proposes the WSNGNL denoising algorithm.First,this algorithm transforms the problem into a subspace denoising problem through the global spectrum low-rank method,which greatly reduces the amount of calculation.Through the decoupling idea of the regular term,the problem is transformed into the denoising problem of each layer of the spectral band of the hyperspectral image projected in the subspace.Through non-local low-rank method,the problem is transformed into a low-rank problem of the non-local matrix.The low-rank approximation method of the weighted Schatten p-norm is introduced to avoid excessive shrink in the low-rank approximation process.We compare the WSN-GNL algorithm with several denoising algorithms between simulated data and real data.Our method is better than other denoising algorithms.There are mixed noises for hyperspectral images,including Gaussian noise,impulse noise,dead-line noise,striper noise,etc.This thesis proposes a novel global non-location low-rank factorization and spare factorization iterative boost method,GNS-ITboost denoising algorithm,which improves the denoising result through iterative regularization.First,GNS-ITboost also uses the global spectral low-rank method for dimensionality reduction.On this basis,we obtain its low-rank factor through the non-local low-rank method,and the non-local low-rank approximation adopts the new non-convex method,the weighted Schatten p-norm method so that it can better approximate the rank function,so as to obtain a better non-local low-rank factor.At the same time,we also propose a new prior knowledge method to estimate the noise of the non-local low-rank part,thereby accelerating the acquisition of better denoising effects.Based on the denoising model constructed by the non-local low-rank factor obtained by the non-convex method and the idea of weighted sparsity,we introduce the proximal alternating minimization method to solve the denoising optimization problem.In experiments with simulated data and real data,our results are better than those of other hyperspectral image denoising algorithms in terms of indicators and visual compare.In this thesis,two algorithms can improve the effect of hyperspectral denoising algorithms,and the non-convex optimization and new optimization frameworks also be introduced.