The Research Of Hyperspectral Image Denoising Based On Subspace Low-rank Representation

Hyperspectral images have widely application in many fields because of the rich spectral information.However,the actual hyperspectral images often contain very complex mixed noise due to the sensors and natural environment,such as Gaussian noise,impulse noise and dead lines,which will not only affect the visual effect,but also decrease the subsequent applications directly.Therefore,the denoising of hyperspectral images is an essential process.The denoising model based on low-rank and sparse matrix decomposition can be established,according to the low-rank property of hyperspectral images.However,this model regards the spectral space of hyperspectral images as a single low-rank subspace.Actually,the spectral vectors can be divided into different categories based on the real object category,which means the spectral space of hyperspectral images should be regarded as the union of multiple low-rank subspaces.In exploring the inner joint low-rank structure of data,the subspace low-rank representation model is very effective,and has been successfully applied for denosing hyperspectral images.On the basis of traditional subspace low-rank representation model,three denoising methods for the removal of mixed noise in hyperspectral image are proposed in this paper,and can be described as follows:(1)Hyperspectral images denoising based on non-local similarity joint subspace low-rank representation(NLSJ-SLRR): The traditional subspace low-rank representation can not make full use of the spatial information in image.Therefore,in the proposed method,the hyperspectral image is firsty divided into many blocks to enhance the local correlation,and the similar blocks are combined by clustering to enhance the utilization of nonlocal similarity in the image.Finally the combined image blocks are denoised by subspace low-rank representation model.A novel dictionary selection strategy is also proposed by preliminary denoising in this method.(2)Hyperspectral images denoising via subspace low-rank representation and spatial-spectral total variation(SLRR-SSTV): In order to further enhance the utilization of spatial-spectral information in hyperspectral images and make it more intuitive in the model,a novel mixed denoising model is proposed by introducing the spatial-spectral total variation into the subspace low-rank representation framework.The model not only can satisfy the low-rank feature of spectral space in hyperspectral images,but also can retain the edge features and enhance the smoothness of both spatial and spectral domain effectively.(3)Hyperspectral images denoising by constrainted subspace low-rank representation and spatial-spectral total variation(CSLRR-SSTV): Dead lines are typical in hyperspectral images,which can be regarded as stripes in extreme cases.In fact,in the hyperspectral image denoising model built by low-rank constraint,it is easy to regard the structured sparse noise,especially the dead lines at the same position in different bands,as the potential low-rank component.Based on the proposed SLRR-SSTV,the low-rank component in the model is strongly constrained to achieve better performance in the removal of sparse noise.In this paper,both simulated and real hyperspectral image datasets are employed to verify the proposed denoising methods,and some of the state-of-the-art methods are applied for comparing.The results of all the experiments show that the methods proposed in this paper can achieve excellent performance both in visual effect evaluation and quantitative indicators.