Research On Super-Resolution For Visible Bands Of Optical Remote Sensing Images In Deep Learning

The spatial resolution of optical remote sensing images is often limited by the sensor resolution and the mechanism of remote sensing image processing,which limits the quantitative level of its analysis and application.The remote sensing super-resolution aims to improve the spatial resolution in terms of software applications.In recent years,the rise of deep learning theories and methods have greatly promoted the development of remote sensing super-resolution.The outstanding problems of remote sensing super-resolution are: firstly,the detection view of remote sensing detection is large,and many targets of interest often occupy few pixels,so the remote sensing super-resolution of full image not only wastes computational resources but also is not conducive to the reconstruction of the targets;secondly,in practical situations,there often lack high-resolution(HR)remote sensing images/videos,hence the supervised based super-resolution methods are infeasible;in addition,for the real remote sensing videos,their kernels and noise are complicated and fully unknown,which brings great difficulties to its high-precision reconstruction.To address these problems,in this dissertation,we conduct research on the local remote sensing image super-resolution,single image and video remote sensing super-resolution in absence of HR target.The main research content is as follows:(1)For local image super-resolution of remote sensing images,a supervised based context guided local super-resolution method is proposed by introducing context information of the local area.The generator of our model extracts the local area features and the context features by a double branch feature extraction module,then train with two discriminators,including a context discriminator and a content discriminator,so as to realize the local area super-resolution of remote sensing images.(2)To address the problem of lack of HR remote sensing image,an unsupervised HR natural image guided remote sensing image super-resolution model is proposed.Firstly,a double branch network structure is designed to super resolve the low-resolution(LR)remote sensing from HR natural image domain and LR remote sensing image domain,respectively.In the natural image driven branch,a super-resolution network with Cycle GAN is constructed to transfer natural images to remote sensing image domain and achieve image reconstruction;and for the remote sensing image driven branch,the model extracts the relationship between different resolutions in the LR images themselves and perform the super-resolution of remote sensing images,which is based on the understanding that the relationship between the resolutions of adjacent images has succession.(3)An unsupervised double branch remote sensing super-resolution model assisted by HR remote sensing images is proposed for the real situations which lack of HR videos as well as the unknown blur kernel and noise in LR videos by using the unpaired HR images and LR videos.The low-scale images/videos in both branches exploit the blur kernel and noise estimated by the LR videos.In the video super-resolution branch,the video superresolution network is trained to learn temporal information based on LR/lower LR video pairs.For the high-resolution image branch the video super-resolution network is trained to extract high-frequency information by constructing video pairs from single high-resolution image,which is prepared by downscaling the HR images to obtain image pairs and then converting them to video pairs by replication.