Hyperspectral Image Denoising Based On Subspace And Low-Rank Learning

Hyperspectral images,which have spatial image with rich structure and tens to hundreds of continuous spectral bands,are usually acquired by imaging the ground objects from both spatial and spectral dimensions through imaging spectrometer located on the different space platforms.Hyperspectral images have a wide range of applications,such as environmental monitoring,military security,mineral mining and food safety detection.Compared to natural images,Hyperspectral images require higher-quality requirements for imaging devices,and their imaging process is more complicated.Although researchers are constantly improving imaging devices and optimizing imaging processes,they are unavoidably corrupted by different types of noise and others interference,which affects the visual quality of the image and reduces the accuracy of image subsequent processing,such as classification,unmixing and target detection.There are many different types of noise in hyperspectral images,including Gaussian noise,impulse noise,stripes and deadlines,and the distribution of different types of noise are different,which brings great challenge for denoising.The main way of hyperspectral image denoising is to make full use of the spatialspectral structure information,design a denoising model that conforms to the characteristics of the data,and then use the learning algorithm to optimize model.Based on it,the main contributions are summarized as follows:1)The current research of hyperspectral image denoising is reviewed from four aspects: sparse representation,matrix decomposition,tensor decomposition and subspace representation.The prior information of hyperspectral image is introduced as well,including spectral band correlation,spatial nonlocal similarity and smoothness,and the sparse,low rank and total variation prior models of hyperspectral image are described.2)A novel denoising method based on subspace superpixel low rank representation and total variation is proposed for hyperspectral image.The method exploits the subspace technique to represent the spectral global low rank of the image,the superpixel low rank representation to mine the spatial local correlation,and the spatial smoothing property design the bandwise total variation prior model.Then the augmented Lagrangian multiplier method is used to optimize the model.Finally,the simulated and real data experiments are undertaken to verify this method can effectively remove the mixed noise.3)A hyperspectral image denoising method based on subspace nonlocal low rank and sparse factorization is proposed.The method considers the spectral correlation of the image,projects the original hyperspectral image into the low dimensional subspace,and then the spatial non-local similarity of image is considered by using non-local low rank factorization.The model is solved by alternating iteration.The effectiveness of the method is demonstrated by simulated and real data experiments.