Research On Hyperspectral Image Denoising Method Based On Deep Learning

As a kind of three-dimensional data with high dimensions,hyperspectral image has a narrow and continuous spectrum,which makes it rich in information and can better characterize the properties of observed ground objects,so it has a wide range of applications.Due to the internal limitation of spectral imaging system and the interference from external factors such as calibration and imaging environment,the hyperspectral images acquired by detection inevitably contain noise.Noise is a bad signal that destroys important,critical information in a clean image.In order to promote further analysis of hyperspectral images,it is necessary to eliminate the noise contained in hyperspectral data before data segmentation,target detection,texture analysis and other tasks.For hyperspectral image denoising problems,for two-dimensional image denoising method can’t flexible adaptive dimensions of hyperspectral data,and will destroy the spatial information of hyperspectral data.The successful removal of hyperspectral noise most of the neural network is supervised learning,the performance of the supervised method depends on a large number of matching noise-clean hyperspectral data samples,so the practical application of this method is limited.There are few unsupervised denoising methods for hyperspectral images,and some unsupervised denoising methods are difficult to guarantee high quality denoising due to low network training efficiency and loss of useful information.Aiming at the above problems,this paper proposes two models for hyperspectral image denoising.The main contents are as follows:(1)An unsupervised network model based on SURE loss function is proposed to solve the problem of supervised neural network dependent pairing training samples.A U-shaped unsupervised neural network is constructed to capture multi-scale features through down-sampling and upsampling,and jump connections are added to transfer shallow features to deeper layers while reducing the loss of data spatial information.To solve the problem of information redundancy caused by jump connection,convolution attention module and nonlocal attention blocks are introduced respectively to suppress redundant information caused by unnecessary features and the transmission of underlying features.A robust network to noise is constructed through the above procedure.To solve the problem of slow and difficult learning mapping relationship in unsupervised network model,Stein unbiased risk estimation mathematical model is introduced.In order to speed up network calculation efficiency and effectively remove spatial and spectral noise,subspace representation method is introduced,and the SURE loss function is constructed based on subspace representation.The network model can find a good mapping relationship about hyperspectral projection matrix through end-to-end learning and mining data intrinsic features.The model outputs hyperspectral projection data and restores clean hyperspectral data by inverse representation of subspace.(2)In the unsupervised network model,aiming at the problem that it is difficult to learn good mapping relations only from low-quality data polluted by noise,a three-layer three-dimensional discrete wavelet convolution neural network is proposed to process high-quality hyperspectral data.Based on a basic U-shaped framework,a U-shaped wavelet neural network is constructed.For feature extraction,using convolution module for pooling operation may result in information loss problem,at the same time in order to ensure accurate data refactoring,the wavelet transform with multi-scale learning ability is embedded in the U-shaped network structure.The feature representation capability is enhanced by pooling operation without changing the main structure.In view of the slow processing of hyperspectral data directly by wavelet network,the feature projection data of clean hyperspectral data is extracted and input into the network by principal component analysis method,and the highquality style feature projection data output by wavelet neural network is fused with the low-quality projection matrix output by unsupervised network.The input of the two branches consists of unpaired input from data sets with different contents.The two network models are combined,and the constructed total loss function is used for training.The high-quality style features are guided to enhance the low-quality data,and the unsupervised network is guided to learn better mapping and recover the hyperspectral data with better denoising effect.