Hyperspectral Image Mixed Noise Removal Based On Low-rank Signal Recovery

Hyperspectral image(HSI)contains hundreds or thousands of narrow spectral bands for the same area on the surface of the Earth.With the wealth of available spectral information,HSI has been widely used in various fields,e.g.,urban planning,precise agriculture,geological mapping,environment monitoring,and so on.Unfortunately,HSI is often contaminated by various types of noise during the process of acquisition and transmission.Furthermore,the noise of HSI appears different statics and intensity in different bands.The noise in HSI not only degrades the quality of the acquired image,but also limits the accuracy of subsequent application.In order to solve the complex noise condition,it is necessary to develop the hyperspectral image denoising method systematically by combining the prior information of the clean hyperspectral remote sensing image and the statistical features of the noise.Aiming at this goal,the paper focused on the research on mixed noise removal of hyperspectral image based on lowrank representation.The research achievements of this dissertation are listed as follows:(1)For the complicated noise condition,we propose a novel hyperspectral image denoising method for comprehensive HSI restoration task by combining the non-local low-rank tensor decomposition and total variation regularization together,here referred as TV-NLRTD.A non-local low-rank tensor is built to simultaneously capture the spatial non-local similarity and high spectral correlation.Furthermore,the spatialspectral total variation(SSTV)regularization is investigated to restore the clean hyperspectral image from the denoised overlapping cubes.The proposed method preserves the image details while enhancing the denoising performance.(2)Since the noise of HSI appears different statics and intensity in different bands,we propose a spectral weighted low-rank matrix approximation model for hyperspectral image denoising,here referred as SWLRMA.The spectral weighted matrix is introduced to balance the data fidelity of the different bands in consideration of their different noise statistics.To further separate the noise from the signal subspaces,weighted nuclear norm minimization is utilized to depict the patch-wise low-rank structure of the high dimensional HSI.(3)The proposed TV-NLRTD model and SWLRMA model are solved by employing an efficient alternating direction method of multipliers(ADMM).Both simulated and real hyperspectral experimental results illustrate the validity and superiority compared with the state-of-the-art HSI denoising algorithm.