Reconstruction And Denoising On Hyperspectral Remote Sensing Image Under CS Framework

In recent years,the hyperspectral remote sensing data with different spatial resolution we are able to obtain has been more and more abundant.However,the process of acquisition and transmission of hyperspectral images unavoidably bring noise pollution.How to obtain more efficient and reliable information from these hyperspectral remote sensing images which are polluted by noise is an important problem to be solved urgently in the current hyperspectral remote sensing research.The theory of compressive sensing breaks through the limitation of Nyquist sampling rate,which can greatly reduce the cost of signal sampling and processing.Therefore,we apply the CS framework to the reconstruction of hyperspectral images and study the denoising algorithm for the reconstructed image deeply.The main research works are as follows.(1)Research on a matching pursuit algorithm based on CS framework.To make up for the defects of the stagewise orthogonal matching pursuit(StOMP)algorithm,this paper presents a variant of StOMP,called backtracking-based and inertia weight index decreasing particle swarm optimization-based StOMP(ba-IWPSO-StOMP)algorithm.As an extension of the StOMP algorithm,in each iteration,the proposed ba-IWPSO-StOMP algorithm uses the IWPSO algorithm to optimize atoms in the measurement matrix,then incorporates a backtracking techinique to select atoms by the second screening.The reconstruction experiments of the classical one-dimensional signal and two-dimensional images demonstrate that,the ba-IWPSO-StOMP algorithm could achieve superior reconstruction accuracy compared with other common OMP-type algorithms.(2)Research on a dictionary learning denoising algorithm based on noise estimation.The K-SVD dictionary learning denoising algorithm has the advantages of short time consuming and good denoising effect,but its application is limited to the condition that the noise intensity of the image must be known.In view of this situation,this paper proposes a method to select the smooth image blocks and combines it with the Singular Value Decomposition(SVD)to achieve the estimation of noise intensity of the image.Then a new denoising algorithm(ABKSVD)which could achieve noise estimation is proposed combining with the obtained noise estimation method and the K-SVD dictionary learning denoising algorithm.The denoising experiments on the classic images demonstrate that the images denoised by ABKSVD algorithm can preserve more details and edge features compared with other two common denoising algorithms.(3)Research on an adaptive denoising algorithm for hyperspectral images combining with feature reduction and dictionary learning.Although the ABKSVD algorithm is suitable for image denoising in spatial domain,it is really difficult to remove the noise in the hyperspectral image effectively only by spatial domain denoising.In view of this situation,a spectral-spatial domain joint denoising algorithm(PCA-ABKSVD)is proposed.Firstly,the principal component analysis(PCA)is applied to process the hyperspectral image to obtain a set of principal component images.Then the proposed ABKSVD algorithm is applied to denoise those principle component images with low energy where noise mainly existed.Finally,the final denoising image was fused according to the corresponding energy of each principal component image.The experiments on simulated hyperspectral remote sensing data demostrate that,compared with several traditional denoising algorithms,the PSNR and the picture quality of the images denoised by the PCA-ABKSVD algorithm are obviously improved.(4)Process the real hyperspectral image.A image processing system for hyperspectral image reconstruction and denoising under CS framework is proposed,which achieving a series of processing for hyperspectral images from sampling to denoising.The experimental results on real hyperspectral images demostrate that the hyperspectral images processed by the proposed system could achieve better results both in reconstruction accuracy and image quality compared with other several processing methods.