Super-resolution Reconstruction Of Remote Sensing Images Based On Compressed Sensing

With the rapid development of society,people have higher and higher requirements for the quality of images,and space remote sensing technology and people's lives have become closer and closer.Opportunities are accompanied by challenges.At the same time that remote sensing images are easy to obtain,due to the limitation of resolution,they cannot meet people's needs.There are various factors in image imaging,transmission,and processing that cause image degradation,resulting in image discrimination.The rate will drop.In order to obtain high-resolution remote sensing images with more high-definition detail information,the emergence of image super-resolution reconstruction technology has changed this situation.It does not change the imaging system equipment,and only processes the image accordingly,thereby improving the resolution of the image.Storage and transmission of massive remote sensing data is also an important factor affecting imaging resolution.Compressed sensing theory can accurately reconstruct the original signal according to the sparseness of the signal,which is extremely helpful for obtaining high-resolution images.This thesis from the perspective of compressed sensing theory,combines traditional learning and deep learning to study the super-resolution reconstruction of remote sensing images.The main research results of this paper are as follows:(1)In image super-resolution,good image prior information is very important for image reconstruction.Non-local similarity and sparseness are two better methods to obtain image prior.In this paper,using this prior knowledge,an image super-resolution reconstruction based on sparseness and low-rank representation of image blocks in a non-local frame is proposed.The framework consists of two steps.The first step is the process of removing noise from a set of matched image blocks,which is described as recovering a low-rank matrix from the noise data.This problem is based on theasymptotic matrix reconstruction model under random matrix theory,so as to obtain parameter less optimal estimation.The second step uses non-local learning sparse representation to suppress artifacts introduced in the estimation.At the same time,a non-local sparse model is used to construct a sparse super-complete dictionary and perform image reconstruction.Experiments show that the reconstructed image can be better.Image texture detail information,effectively suppressing noise.(2)It is proposed to apply Markov random fields to image super-resolution reconstruction,and use the training set images to construct a database during the training phase.The low-resolution image blocks and high-resolution image blocks in the database are simulated by Markov networks.At the same time,the image is decomposed into structural parts and texture parts,because the main difference between low-resolution and high-resolution images lies in the high-frequency information part,that is,the texture part.Perform dictionary learning on the texture part of the image,use the bicubic interpolation method to enlarge the color information of the image and the structural part of the image,merge the reconstructed image and the image obtained from the bicubic interpolation to get the final result.It is proved by experiments that the quality of the image reconstruction is both subjective visual effects and objective indicators achieved good results.(3)Block-based compression is a lightweight compression method,which is mainly suitable for processing high-dimensional images and videos,running on local blocks,using low-complexity reconstruction operators,and requiring less memory to store the sensing matrix.In this paper,a block-based deep learning algorithm is proposed.This algorithm uses fully connected networks for linear perception and nonlinear reconstruction of blocks.During the training phase,the perception matrix and nonlinear reconstruction operators are jointly optimized.And calculation time have obvious advantages.