Research And Application On Dictionary Learning Method For Remote Sensing Image Restoration

The remote sensing technology,which is able to produce various types of the earth observation data in a dynamic,rapid and accurate manner,has become the main information source of geographic information system and the main method to update the data.However,caused by remote sensing link,the distortion of the image radiation quality limits the precision for the extraction and matching of the feature points,which further limits the evaluation results of remote sensing image geometric positioning accuracy.Therefore,in this thesis,we research on the dictionary learning algorithm for remote sensing image denoising,deblurring and super-resolution.Then we integrate the proposed restoration algorithms on the extraction and matching of feature points to improve the accuracy of the geometric positioning evaluation.The main work of this paper can be concluded as follows:(1)A multi-scale fractional-order dictionary learning(MFDL)model is proposed for image denoising.To solve the deviation problem of both the sparse coefficients and dictionary in the conventional sparse representation based algorithms,we introduce the fractional-order technique into the dictionary learning method.In the multi-scale image domain,we re-estimate the singular value metrics of the noisy image and construct the fractional-order sample space.Then we utilize the sparse coding and dictionary updating to suppress the deviation of both the sparse coefficients and dictionary.Experimental results demonstrate that the MFDL model improves the accuracy the sparse coefficients and effectively overcome the impact of noise.(2)A novel non-local tensor dictionary learning model is proposed for multi-spectral image deblurring.We exploit nonlocal structural self-similarity of core tensors,and incorporate it into the tensor dictionary learning algorithm as a prior regularization term.First,making full advantages of spatial similarity and band correlation,three-order tensor is used to represent the image patches of multi-spectral remote sensing image,and the core tensors are constructed through high-order singular value decompositions.Then,we calculate the novel the core tensors by utilizing the non-local structural similarity of tensors.Finally,the blurred multi-spectral remote sensing images are reconstructed using the tensor dictionary learning algorithm.The experimental results demonstrate that the proposed model improves the accuracy of core tensor and enhances the restoration effect of multi-spectral images.(3)By modeling the core tensor as a Laplace distribution with non-zero mean,a novel adaptive tensor dictionary learning model is proposed for multi-spectral image super-resolution.First,we calculate core tensors of multi-spectral remote sensing images by high-order singular value decompositions.With the prior panchromatic images,both the mean and variance of the Laplacian distribution is then adaptively estimated,and then the core tensors are reconstructed.Finally,we construct the super-resolution images using the tensor-based dictionary learning algorithm.Experimental results show that the proposed model can accurately estimate the mean and variance of the Laplace model,which helps to preserve the spectral fidelity and improve the performance of the super-resolution algorithm.(4)In order to improve the accuracy of the geometric positioning evaluation model,the proposed restoration algorithms are integrated into a unified framework,and a novel feature point extraction model is proposed based on the sparse classification of remote sensing radiation features,which can be treated as a two-stage strategy.In the first stage,based on the initial feature points,we establish the geometric relationship between the remote sensing images and their corresponding multi-resource reference images,and compensate geometric deflection for the evaluated images.On the second stage,in order to solve the degradation of the image(e.g.,noisy panchromatic images,blurred and down-sampled multi-spectral images),we reconstruct the remote sensing image by the proposed dictionary learning models to provide abundant textures for feature points extraction and matching.Then,we cluster the sub-regions of images using sparse classification based on remote sensing radiation features.Finally,the faithful geometric positioning evaluation results are abstained based on the feature points extracted by utilizing adaptive SURF algorithm.The experimental results demonstrate that the proposed model can extract the feature points with high accuracy and uniform distribution,which leads to the accurate geometric positioning evaluation results.