Low-rank Modeling And Regularization Method For Hyperspectral Image Denoisin

Hyperspectral imaging is an important mean of remote sensing detection.Hyperspectral images are widely used in geological prospecting,military investigation,environmental monitoring and other fields because of their rich spatial and spectral information.Affected by factors such as hardware instruments and atmospheric interference,hyperspectral images will suffer serious noise pollution during the collection and transmission process,including Gaussian noise,band noise and dead lines,etc.,which will reduce image quality and affect the subsequent application of hyperspectral images application.Hyperspectral image denoising is of great significance as a preprocessing operation for subsequent applications of images.Hyperspectral image denoising belongs to image restoration,and its purpose is to recover a noise-free image from the observed image,which is usually an ill-conditioned problem.How to construct a suitable image prior and design an optimization model based on regularization is an important issue in image denoising.Based on this,the main research contents of this article are as follows:（1）Research a kind of multi-channel image modeling and denoising mechanism with total variation of the collaborative norm.Stacking gradient images in different directions of the multi-channel image can form a three-dimensional matrix,and use different L_p,q,r norms to constrain each dimension of the matrix in turn,and different denoising models can be obtained.Through the denoising experiment of color image and multispectral image,the denoising effect of using different collaborative norms for different spectral channels is analyzed.The results show that the selection of the collaborative norms L_p,q,r is related to the correlation between the image color channels.（2）Propose a hyperspectral image denoising algorithm based on matrix low rank and collaborative total variation constraints.Since the observed hyperspectral image contains noises such as Gaussian,salt and pepper,impulse and dead line,the paper models the noisy hyperspectral image as a degraded model composed of latent hyperspectral image,Gaussian noise component and sparse noise component.Aiming at the space-spectrum joint structure of the image,this paper expands the hyperspectral data into a plane matrix,and makes full use of the row sparsity characteristics of the gradient of the expanded matrix to construct a collaborative total variation regularization term to describe the spatial local smoothness of the image,using the matrix kernel norm characterizing the low-rank structure characteristics of the expanded matrix,and an optimization model of the joint matrix low-rank and collaborative total variation is established.and finally the model is solved by the Alternating Direction Method of Multipliers（ADMM）.The experimental results show that the proposed method has good edge retention ability and is superior to the series of benchmark regularization denoising methods in removing mixed noise.（3）Propose a hyperspectral image denoising algorithm based on low-rank tensor and space-spectral collaborative norm.The matrix-based denoising model will destroy the original three-dimensional structures of the hyperspectral image.At the same time,the gradient image of the hyperspectral image in the spatial dimension and the spectral dimension conforms to different statistical distributions.In response to the above problems,this paper constructs a new L_2,1,1 collaborative norm constrained space-spectrum total variation regularization term to fit the difference in the statistical distribution of gradient images of different dimensions and explore the image space more fully local smoothness between and spectra.In addition,tensor low-rank can better mine the internal correlation of hyperspectral images in the spatial and spectral dimensions,and fully retain the low-rank structural information of the image.Combining the tensor low-rank prior and the space-spectral collaborative norm,this paper proposes an optimization model for hyperspectral image denoising,and designs a high-performance solution algorithm under the framework of alternate direction iteration.Experimental results show that compared with matrix low-rank denoising algorithms,this algorithm further improves the denoising performance,effectively removing mixed noise while maintaining the space-spectrum joint characteristics.