Study On Super-resolution Reconstruction Of Optical Staring-imaging Satellite Images

Staring remote sensing is a new remote sensing technology equipped with a 2-D staring imager.Compared with classic pushbroom and whiskbroom imaging,staring imaging has the advantages of longer integration time and higher time resolution.Via pointing control,staring remote sensing satellite can capture multiple images of the same scene during a short time period,which benefits automatic object detection and tracking,and enables new image processing capabilities.Image super-resolution(SR)reconstruction is an effective way to generate a high-resolution(HR)image from multiple low-resolution(LR)images which are aliased as well as shifted with subpixel precision.When the image is undersampled,the high frequency is aliased with the low one.The SR technology can recover the high frequency from multiple LR images without changing the exsiting systems,thus improve the saptial resolution and sharpness without high expense.This thesis studies the key technologies of the SR of the staring remote sensing images through theoretical analysis,simulation and real images verificaiton.The purpose of this paper is to convert the high time resolution to a higher spatial resolution which is important in improving the accuracy of object detection and location.The research work in this thesis mainly focuses on building the observation model,spatial and photometric registration,PSF estimation and SR based on the maximum a posteriori(MAP),and online SR.Based on the need of SR,this thesis analyzes the image formation process and establishes an accurate imaging model.In this thesis,the transmission and scaterring process of the radiaton are analyzed,resulting in a linear intensity variation model which inludes the contrast gain and intensity bias;the blurring factors are also analyzed,including atmospheric scattering,relative motion,diffraction of optical system and finite size of detector;then the geometric model of the detector is deduced under some assumptions.All the degrading factors are incorporated into the imaging model which describes the formation of LR images from the HR scene.A jointly image registration method both in spatial and intensity domain is proposed to meet the SR algorithms' requirements of high registration accuracy.Firstly,an intensity invariant feature detector named SURF is used to estiamte motion parameters.Then photometric parameters are directly estimated based on RANSAC.Lastly,these parameters are refined alternatively by the correlation-based registration method using optimization method.To improve the accuracy and robustness,an inlier mask is introduced to eleminate the outliers in the correlation-based registration method.Experimental results show that the proposed method has a good performance in robustness and registration accuracy.A pipeline of SR algorithm based on MAP is proposed.Firstly,in the LR image preprocessing stage,the region of interest(ROI)is extracted using image template mathcing based on SURF,and the low-quality images are removed using the sum of Sobel gredients.Secondly,the motion and intensity paramters are estimated by the proposed image registration method.Then the point spread functions(PSFs)are estimated by a blind deconvolution method in which the PSFs are reparameterized by Gaussian functions.The last step is to reconstruct an HR image by the image fusion based on the MAP method,and an edge-preserved image prior named Huber prior is combined to suppress the artifacts and noise.To find the global optimal solution,image registration parameters and PSFs are updated alternatively along with the HR image.The influence of each parameter is studied by simulation,and some meaningful conclusions are obtained.The experiments on real images show that the proposed SR method improves the sharpness and spatial resolution,and reduces the noise effectively.A quick SR algorithm working in online mode is proposed.In order to find the global optimal solution,the SR algorithms generally work in batch mode,i.e.,all LR images are processed simultaneously.While in the online SR algorithm,an LR image is used to update the current HR image immediately when captured.In order to improve the robustness and accessibility,the step size of the steepest descent method is modified by a noise-related parameter.Since the proposed method only needs to save the first and the current LR image,the current HR image,and only one LR image is processed at a time,it has advantages of fast computation and saving memory resources.Experimental results show that the performance of online SR algorithm is not as good as MAP-based method,but it is much superior in computation speed,thus has the prospect in real-time SR.