Hyperspectral Image Denoising With Bilinear Low Rank Matrix Factorization

Due to the physical limitations of hyperspectral imaging sensors,interference and photon effect in the process of image data transmission,the actual hyperspectral images are often inevitably interfered by different types of noise.The existence of noise not only affects the visual effect of the image,but also has a negative impact on the interpretation task and quantitative product inversion of the remote sensing image.Therefore,how to design the algorithm to remove the noise of hyperspectral remote sensing image and improve the visual quality and effective utilization efficiency of image data is a hot issue in the field of hyperspectral remote sensing image processing.In this paper,the bilinear low rank matrix factorization model is proposed to solve the problem of mixed noise removal of hyperspectral images,which combines the spatial low rank and spectral low rank characteristics of hyperspectral remote sensing images under the regularization framework.The main work and innovations are as follows:1)The bilinear low rank matrix factorization model based on bi-nuclear quasi-norm is proposed.The existing low rank priors can not describe the spatial low rank and spectral low rank characteristics of hyperspectral remote sensing image at the same time.To solve this problem,this paper proposes the bi-nuclear quasi-norm which can describe the spatial low rank feature and the spectral low rank feature of hyperspectral remote sensing image simultaneously.The bi-nuclear quasi-norm is also closer to the rank function.Based on the bi-nuclear quasi-norm,the bilinear low rank matrix factorization model for noise removal of hyperspectral images is proposed.Because the model can well describe the low rank characteristics of hyperspectral image,it can not only remove the mixed noise,but also keep the information of the image.2)The bilinear low rank matrix factorization total variation model is proposed.The bilinear low rank matrix factorization model only considers the spatio-spectral redundancy of the whole hyperspectral remote sensing image,but ignores the characterization of the local structure of the image,such as the segmented smooth structure of the image.In order to solve this problem,this paper proposes a bilinear low rank matrix factorization total variation model by combining the anisotropic space spectrum total variation priori and bilinear low rank matrix factorization model.The model not only constrains the spatio-spectral low rank feature of hyperspectral remote sensing image from the overall point of view,but also constrains the structural feature of image from the local point of view,improving the denoising ability of bilinear low rank matrix factorization model.