The Research Of Optical Remote Sensing Image Blind Restoration Based On Sparse And Low Rank Model

Optical remote sensing is an important means of earth observation,which is widely used in many fields,such as urban planning,military investigation,environmental monitoring.However,the remote sensing image quality will degrade due to the attitude of satellite platform,detector response and atmospheric turbulence in the process of remote sensing imaging.In order to obtain high quality remote sensing images,the degraded images need to be resorted.Most of the existing image restoration algorithms are assumed that the point spread function and the noises are known.In the imaging process,the degradation factors of remote sensing images are complex;it unable to get the point spread function and noise distribution of remote sensing image accurately.It is very necessary to restore image in the case of the degradation factors are unknown.This thesis focuses on the problem of optical remote sensing image quality improvement.The image blind restoration is based on sparse and low rank model,it include image blind denoising,blind deblurring and Hyperspectral remote sensing image denoising.The main contents and innovation of the paper are as follows:The noise distribution of remote sensing image is analyzed and modeled by Dir-GMM;it can effectively describe the complex noises of remote sensing image.The advantage of the proposed noise model is that the Dirichlet process mixture model can automatically determine the number of components and the ratio of the mixing components in the Gauss mixture model.Traditional matrix decomposition methods assumed that the noise terms satisfy Gauss distribution and the rank of matrix is needed to be specified in advance.The traditional low rank matrix decomposition method can only denoise the Gauss noise and the specified rank is not necessarily optimal,making the denoising performance worse.In order to reduce the complex noise in remote sensing images using low rank matrix decomposition method,the traditional low rank matrix decomposition model is improved by the Dir-GMM noise model and described in nonparametric Bayesian framework.The automatic correlation model is used to determine the rank for automatic selection model.The proposed Dir-GMM low rank matrix decomposition model is used for blind remote sensing image denoising,it don't need to know the noise type and noise intensity in advance,and the rank of the matrix also not required.Based on the influence of the blur operation on image sparseness and low rank priors,a new algorithm for blind debluring of remote sensing images with image and gradient combined low rank priors is proposed in this paper,and the weighted nuclear norm minimization method to further enhance the effectiveness of low-rank prior for image deblurring,by retaining the dominant edges and eliminating fine texture and slight edges in intermediate images,allowing for better kernel estimation.In the blur kernel estimation step,the multi-scale strategy is used to optimize the blur kernel from coarser to fine.A weighted nuclear norm minimization method to further enhance the effectiveness of low-rank prior for image deblurring,by retaining the dominant edges and eliminating fine texture and slight edges in intermediate images,allowing for better kernel estimation.After obtaining the best estimated blur kernel,the deblurring image is obtained using non blind deblurring methodA hyperspectral image(HSI)is considering as 3 order tensor,the hyperspectral image denoising can be regarded as low rank tensor decomposition problem.The observed hyperspectral noise image is decomposed into low rank tensors and sparse tensors to denote the denoised hyperspectral image and noise respectively.The low rank tensor decomposition problem can be regarded as the expansion of low rank matrix decomposition in the high dimension,and described under the Bayesian framework.A hierarchical Bayesian low rank tensor decomposition model is proposed for hyperspectral remote sensing image denoising.The model is solved by nonparametric Bayesian framework,and the observed data are decomposed into low rank tensor and sparse tensor,and the optimal rank of tensor can be adaptively determined.In order to make full use of the Nonlocal Similarity prior,first,a HSI is divided into small tensor blocks.Second,similar blocks are gathered into clusters,and then a low tensor decomposition model is constructed based on every cluster for denoising.Due to the limitation of imaging equipment,it is hard to obtain high spatial resolution and high spectral resolution remote sensing images.In order to solve the inherent contradiction between spatial resolution and spectral resolution of remote sensing images,a non negative structure sparse representation method is proposed for high spatial resolution and high spectral resolution remote sensing image fusion.The real spatial response and spectral response are unknown in remote sensing image fusion,and an efficient method is proposed to estimate the spatial and spectral response.