An Unsupervised Deep Learning Method For Hyperspectral Image Denoising

With the maturity of earth observation technology,hyperspectral images have been widely used in resource exploration,target recognition,urban planning,environmental monitoring,and other fields.However,limited by the conditions of imaging and equip-ment,the acquired hyperspectral images are inevitably polluted by various noises,such as Gaussian,salt and pepper,bands,bad lines,Poisson noise,etc.The existence of noise not only greatly reduces the visual quality of the image,but also seriously affects its sub-sequent applications.Hyperspectral image denoising aims to estimate and invert clear and high-quality hyperspectral images from noise-contaminated hyperspectral images.It is one of the important topics with crucial research significance and extensive applica-tion value to be solved urgently in the interdisciplinary of mathematics and information.The traditional model-based method uses image domain knowledge to establish an op-timization model to solve the problem of hyperspectral image denoising,which has the problem of insufficient prior information expression ability.For this reason,a large num-ber of deep learning-based methods have emerged in the field of remote sensing image denoising.Recently,unsupervised deep learning methods have attracted the attention of researchers worldwide due to its advantage of only requiring observation data,but they face the difficulty of image reconstruction in complex noise scenes.In response to the above problems,this paper proposes an efficient,high-performance and robust hyperspec-tral image denoising model and designs a corresponding solution algorithm by combining the respective advantages of deep learning and traditional model methods.The research content of this thesis is as follows:1.At present,based on unsupervised learning,such as deep image prior(deep image prior,DIP^[1]),the method proposed by Bao et al.^[2],etc.,although much progress has been made in the field of denoising.However,the performance degrades severely in face with complex noise scenarios（such as pulses and strip disturbances）.To this end,this paper proposes a new hyperspectral image denoising model based on unsupervised deep learning,called the Latent Subspace（LSS）model.2.Hyperspectral images have two characteristics of large data scale and complex noise structure,which are difficult to process.This paper proposes a subspace represen-tation to reduce computational cost and process images more efficiently by projecting hyperspectral images into low-dimensional subspaces.Finally,we cleverly integrate the proposed model with the proposed subspace representation,which can effectively and efficiently deal with the complicated noise removal problem of hyperspectral images.3.We design an efficient algorithm for solving the proposed model based on the method of alternating direction multipliers,and verify the numerical convergence of the algorithm through experiments.The denoising performance and speed advantages of the proposed method are demonstrated through the evaluation protocols in our experiments.Moreover,the visual results of the reconstructed images intuitively demonstrate the supe-riority of our algorithm in preserving local details.