Research On Hyperspectral Image Denoising Method Based On Dual Subspace Learning Method Combined With Graph Embedding Constraints

Hyperspectral remote sensing images have been widely used in terrain classification,mineral exploration,environmental monitoring,military surveillance and other application fields due to their large amount of spatial information and spectral information.However,the hyperspectral image will inevitably be damaged by the mixture of various noises during the acquisition process,which greatly reduces the discriminative performance of the hyperspectral image,and then directly affects the analysis and application of the image.Therefore,hyperspectral image denoising is an important and challenging research area for further image analysis.In the past few decades,hyperspectral image denoising techniques have developed rapidly.However,past denoising methods often ignore the fact that pixels in hyperspectral images can be decomposed into low-rank subspaces and the spatial structure of hyperspectral images is self-similar.In order to improve the effectiveness and robustness of hyperspectral image denoising under complex noise conditions,we make full use of the prior information of hyperspectral images,and propose two denoising methods.The main work is as follows:1)Hyperspectral images have hundreds of spectral dimensions.Based on the spectral band correlation,we propose a hyperspectral image denoising method based on subspace combined with graph embedding constraints.First,a subspace representation of the hyperspectral image is performed,and then a graph embedding constraint is imposed on the subspace coefficient matrix using spectral correlation.Iterative optimization is then performed using the alternating direction method of multipliers.Finally,the superiority of this method in removing mixed noise is verified by simulation experiments and real experiments.2)Considering the spatial and spectral properties of hyperspectral images,a dual-subspace image denoising algorithm based on spatial and spectral information is proposed.Decompose hyperspectral images into spatial and spectral subspaces using non-negative matrix factorization and graph embedding constraints.Spatially,similar blocks in space have similar spatial subspace coefficients through graph embedding constraints.Spectrally,similar blocks in adjacent bands have similar spectral subspace coefficients through graph embedding constraints.Then,the model is optimized using alternate iterations.Finally,the experiment proves the effectiveness of this method in removing noise.