Algorithm Research On Super-resolution Reconstruction Of Optical Remote Sensing Images

There is an urgent demand for high resolution image in the field of surveying and mapping,military and civilian areas.How to obtain high resolution remote sensing image is one of the most important research fields for remote sensing researchers.However,in the acquisition process,the image can be affected by various degradation factors,which makes the image can't fully meet application requirements.Super resolution reconstruction technique uses signal processing knowledge,basing on the existing imaging system,reconstructs high resolution image with rich details and high pixel density from single frame image or multiple frame low resolution images.Super resolution reconstruction has been widely used in many fields,which becomes to be one of the most promising research directions in the field of image restoration.This dissertation first introduces the research significance,the present situation and development trend of the research subject.And then describes the theoretical basis of super-resolution reconstruction,the classical algorithms categories and theoretical framework,remote sensing imaging technology and reconstruction evaluation method.Accurate sub-pixel registration method and its parallel design scheme have been studied based on feature point detection theory.Basing on spatial regularization,the main clue of this thesis,combing with application of partial differential equation and sparse representation learning theory,single frame and multi-frame super-resolution reconstruction algorithms have been proposed and realized.This paper put forward a variety of super-resolution reconstruction algorithms,and has made certain theoretical innovation.More researches applied to linear array and array images in the field of remote sensing have been studied on the basis of existing proposed algorithms.The research results and innovation of this paper include the following aspects:1.Multi-frame image registration algorithm and its parallel design scheme have been proposed.The algorithm implements new feature point detection method based on an improved adaptive gradient bilateral tensor filter.More accurate angular position information has been obtained by sub-pixel fitting of weighted Gauss surface,which better ensures the registration accuracy.Finally,the algorithm is extended into multi scales description and feature point descriptors are generated to implement registration.In addition,the proposed method is designed for parallel analysis in operation parallel level and task parallel level based on CUDA on GPU platform and OpenMP on multi-core CPU platform.2.A new multi-frame super-resolution reconstruction algorithm has been proposed,which combines two regularization terms including the nonlinear anisotropic diffusion tensor regularization and the improved continuity constraint of gradient vector flow regularization.The algorithm is designed based on partial differential equation,constructs regularization item by nonlinear diffusion tensor,making full use of its directional selective smoothing characteristics,and more intuitively describes the filter performance in flat,edge and corner structure regions.During the reconstruction solving process,the optimization process solves along tangential direction on edge and texture region,and limits its optimization in gradient direction.In addition,the improved gradient vector flow field uniform continuity regularization term can better restrict high and low resolution images,which shows great robust to noise and better keeps texture structure information.3.An improved K-PCA adaptive dictionary learning algorithm which improves the initial clustering center has been proposed.Combing with nonlocal nonlinear tensor filter regularization and nonlinear diffusion edge preserving kernel regression regularization,an improved single frame super-resolution reconstruction method based on learning theory has been implemented.First,measure the initial cluster centers basing on structure tensor space property measurement.Then,consider the similarity measures of gray and geometric structure,combing with nonlinear tensor,bilateral tensor and fuzzy entropy,to put forward an improved regularization term basing on nonlocal means algorithm.Besides,the second regularization term basing on nonlinear diffusion tensor kernel regression and Gaman-McClure kernel function is proposed.Furthermore,simulation experiments analysis have been done,and the initial inputs of multi-frame reconstruction are substituted by single frame reconstruction results to analyze multi-frame reconstruction application.4.A unified framework based on partial differential equation for single frame super-resolution reconstruction algorithm has been proposed.The proposed algorithm fully combines with local edge structure and texture feature analysis,high order tensor analysis,multiple orientations estimation,gradient vector flow field and nonlinear diffusion tensor theory,which can better adapt to various noise levels of single image reconstruction.The anisotropic nonlinear structure tensor considers different channel information to prevent singular results,which preserves edge texture and get better de-noise effect.Junction,corner and structure information can be more accurately estimated by high order structure tensor.As a unique descriptor of orientation,mixed orientation parameter(MOP)can be separated into two orientations by finding roots of a second-order polynomial in the nonlinear part.The improved shock filter can better realize the edge enhancement in noise situation.5.On the basis of researches of parallel sub-pixel registration algorithm,multi-frame and single frame reconstruction algorithms based on partial differential equation,and single frame reconstruction algorithm based on improved sparse representation,all the proposed algorithms have been applied to real line-scanning data and area array camera data.Furthermore,the single frame reconstruction result is used as the initial input for multi-frame reconstruction simulation verification.The experimental results show that the improved feature point registration algorithm designed in parallel scheme achieves more accurate registration and speeds up registration process for remote sensing images.Reconstruction algorithms based on partial differential equation and machine learning theory achieve better reconstruction results for remote sensing images both subjectively and objectively.