Research On Super-resolution Reconstruction Of Dynamic Sequence Remote Sensing Images

With the constant progress in remote sensing satellites towards miniaturization and commercialization,improving the resolution of the dynamic sequence remote sensing data by information processing technology evidently has the practical significance and research value.Therefore,the super-resolution reconstruction technology of remote sensing image has received extensive attention,and a lot of research progress has been made in recent years.Sequence reconstruction algorithm is the main research content of remote sensing image in super-resolution technology.Meanwhile,the boom of machine learning in recent years makes the learning-based super-resolution algorithms develope greatly.this paper mainly focuses on the array remote sensing image,systematically analyzes its characteristics in each part of reconstruction algorithm,and studies it with reconstruction and learning algorithm.This paper focuses on dynamic sequence remote sensing image,and the resolution amplification limit,preprocessing algorithm,characteristics in super-resolution algorithm,evaluation criteria and resolution magnification estimation method are systematically analyzed.Meanwhile,the reconstruction and learning algorithm are combined in a comprehensive study.Firstly,the paper explains the background and significance of this subject,investigates the domestic and foreign research status,and summarizes the development trend of super-resolution reconstruction technology for dynamic sequence remote sensing images.Secondly,the degradation process of remote sensing image is described,the theoretical basis is discussed,the applicability of super-resolution algorithm based on reconstruction and learning to sequence remote sensing image is analyzed,the quality evaluation standard of super-resolution image is summarized,and the resolution magnification estimation method based on target recognition is proposed.Thirdly,the amplification limit of remote sensing image is analyzed and calculated,and the methods for selecting the quantity and quality of low-resolution images are proposed.Fourthly,the preprocessing and denoising methods are proposed for the poor-quality remote sensing images,and the sequences of reconstruction and denoising procedure are analyzed in detail.Fifthly,the detection methods of sub-pixel points are summarized,an interpolation method of average coverage registration that can restore more real high-frequency information is proposed,and aiming at the matching problem of different brightness,the multi-brightness layer matching method is proposed.Finally,every aspect of super-resolution reconstruction on the final results are analyzed,and a combination of learning algorithm and reconstruction algorithm is studied.According to these research contents,the innovation results mainly include:1.The resolution magnification algorithm based on target recognition method is proposed.According to a large number of buildings in more remote sensing image and different resolution with different recognizing ability,using bilateral edge detection algorithm to statistics the number of buildings can estimate the resolution magnification by their specific value.2.On the basis of the super-resolution reconstruction amplification limit perturbation theory,the characteristic parameters of remote sensing images are added to calculate the amplification limit of real remote sensing images.A method for selecting the appropriate number of LR images based on noise and image gradient is proposed,and a table is made for researchers to select conveniently.For real remote sensing images,the signal-to-noise ratio(SNR)algorithm is improved and used to judge the LR image quality.By analyzing the frame relation of the dynamic sequence remote sensing image,the method that satisfies the sub-pixel displacement in LR image is defined.3.A self-estimation block matching 3D denoising algorithm(SE-BM3D)is improved to pre-process LR remote sensing images which do not meet the quality requirements.This algorithm adds the model of modal noise and gaussian noise that are more consistent with remote sensing image on the basis of BM3 D,and uses the self-estimation of noise power density spectrum of the obvious uniform background in the remote sensing image to increase the accuracy of noise removal,so as to retain more information details.Experimental results show that the SE-BM3 D algorithm can effectively improve the SNR value of LR images to meet the requirements,and maintain details better than existing denoising methods.As a preprocessing denoising algorithm,it can obtain more clear super-resolution results than other algorithms.Aiming to the LR images which can not use the SE-BM3 D algorithm or fulfill the quality requirements of the reconstruction images because of large noise,the denoising and reconstraction order problems are solved.Meanwhile,according to the LR image gradient values and the noise standard deviation,the strategies and frameworks of super-resolution reconstruction are provided.4.Image matching is a significant part in super-resolution reconstruction.Based on remote sensing imaging process,the average coverage of interpolation algorithm considers the reduced sampling interpolation algorithm,but also takes the blur and integral into account.Thus,this method has the ability of restoreing more real high-frequency information.On the other hand,the uneven and dramatic illumination variations between images can impact the matching result seriously.Therefore,this paper presents a method to match features efficiently under uneven and dramatic illumination changes.This method extracts and describes illumination-invariant interesting points from matched multi-brightness layers which are obtained by a set of contrast stretching functions and prior information based on original images.Layers matching is insensitive to large unevenness of illumination changes and provides similar images in brightness and structure,so that the effects of large uneven illumination changes can be reduced greatly.This algorithm is compatible with most detectors and descriptors.To accelerate the computing speed,the features from accelerated segment test(FAST)detector and improved speeded up robust features(SURF)descriptor are chosen in this paper.In addition,the combination of priority Hamming distance matching and Lowe's matching algorithm is first proposed to increase the matching speed.This method is generic and can be used in most point matching under all varying illumination conditions.Experimental results show that the proposed algorithm is better than the existing OSID,IRFET,SIFT and SURF algorithms in matching the brightness changes,and has the rotation and scale invariance required by the excellent feature point operator.5.A simultaneous super-resolution reconstruction method based on dynamic sequence image is proposed.Firstly,the algorithm can obtain more accurate anisotropic blur matrix based on the kown attitude angle of imaging system geometry model to improve the precision of the degradation model.Secondly,through the analysis of every aspec such as blur core shape & standard deviation,matching matrix and the prior parameters on the influence to the reconstruction result,the discovery that they are not independent effect on the final result is found.For getting better HR image,it is not only estimating all aspects as accurately as possible,solving the problems of the cooperation between parameters is the more important at the same time.Thus,the ideas of cross validation from machine learning is introduced,and the cross-validation simultaneous optimization super-resolution algorithm is proposed.This method optimizes the prior parameters,HR image,matching matrix and blur core standard deviation parallelly,and uses cross validation to test at the same time.Through this algorithm,the global matching error caused by the inherent noise of the imaging system is compensated,the inaccuracy of initial value estimation of each parameter is taken into account,and the coordination between parameters is strengthened.The results of simulated image and real remote sensing image experiments show that the algorithm in this paper can effectively improve the sharpness and resolution magnification rate of remote sensing image.Except for the specific anisotropic blur estimation,the algorithm is also universal for other types of images.