Research Of Hyperspectral Image Denoising Algorithm Based On Tensor Decomposition

As a special form of data,noise appears with the appearance of signal.According to its probability density function,it can be classified as Gaussian noise,salt and pepper noise,gamma noise,Rayleigh noise and so on.The generation of image noise will disturb the information contained in it,which has a great impact on the subsequent image data processing,such as data mining,machine learning,image recognition.Classical image denoising algorithms for two-dimensional image data,the application of a variety of prior knowledge and mathematical models,in the premise of preserving the details of the original image to remove the impact of noise to the greatest extent.However,for the special information such as hyperspectral image,because of its large dimension of spectral dimension,if the image data is expanded into two-dimensional matrix data along the spectral dimension and processed by the classical denoising algorithm,the relationship between spectral dimension and spatial domain will be lost.In recent years,researchers have proposed a variety of denoising algorithms based on tensor decomposition model to keep this connection.This paper focuses on the research of hyperspectral image denoising algorithm based on tensor decomposition:1.A nonlocal tensor low rank decomposition denoising algorithm is proposedThe objects of the same material in hyperspectral images have the same spectral performance in the spectral band,and they have rich nonlocal similar information.Taking advantage of this feature of hyperspectral image,this method uses nonlocal spatial pixel information to complete the pixels to be estimated affected by salt and pepper noise.After completion of the completion,the fourth-order tensors aggregated from similar blocks are denoised by using sequence truncated high-order singular value decomposition(st-hosvd).2.A low rank tensor denoising model with t-gamma norm as low rank regularization term is proposed.In classical low rank tensor denoising models,kernel norm is often used as the low rank regularization term,and the kernel norm constrains the singular values of different sizes equally.In order to solve this problem,gamma function,which is currently used for matrix low rank recovery,is introduced into the mixed noise denoising model of high-order data,and t-gamma norm is used as the low rank regular term.In order to remove mixed noise better,salt and pepper noise and Gaussian noise are modeled respectively,and constrained by l₁ norm and Frobenius norm.Finally,the alternating direction multiplier method(ADMM)is used to solve this problem.3.A nonlocal low rank tensor denoising algorithm is proposed.Combined with the nonlocal similarity features of hyperspectral image,the hyperspectral image is segmented and aggregated into a fourth-order tensor.In order to deal with this kind of high-order tensor data,combining gamma function with Tucker decomposition model,Tucker gamma norm,which is more general than t-gamma norm,is proposed as the low rank constraint term of low rank denoising model.Low rank recovery is performed for each aggregated fourth-order tensor block,and denoising is completed in the process of model solving.The final hyperspectral image is synthesized by weighting the denoised tensor blocks.In this paper,the peak signal-to-noise ratio and structural similarity are used to objectively evaluate the three tensor denoising algorithms.Compared with the existing classical hyperspectral noise filtering methods,the effectiveness of the algorithm in noise filtering is verified.