The Research And Application Of Satellite Imagery Super-resolution Based On Multi-scale Residual Deep Neural Network

Remote sensing satellite is a rapidly developing and high-tech industry,which observe objects on the ground from outer space,is widely used in various real applications,such as environmental monitoring,resource exploration,disaster warning,and military applications.However,the real scene remote sensing satellite images are affected by optical components,imaging noise,and motion blur.The resolution of images from remote sensing satellites cannot meet the requirements of real satellite image.Super-resolution(SR)technology can overcome hardware and technical limitations,improve the spatial resolution of images through software.Hence,SR reconstruction can improve the efficiency of remote sensing satellites economically and effectively.Learning-based satellites imagery SR algorithms utilize the prior information provided by training dataset to predict the missing high-frequency information in the low-resolution image and obtain the relatively fine performance.Deep learning-based algorithms are the most popular in this field.Deep learning provides a powerful super-resolution solution that simulates nonlinear expression,greatly improving the performance in both subjective and objective.This thesis focuses on the restoration of high-frequency information in image reconstruction and the multi-scale characteristics of satellite imagery,and launches research on the convolutional neural networks with "deeper layers" and "wider networks".The main research results are as follows:1.Satellite image super-resolution via relay residual network,the information gap between low-resolution(LR)images and high-resolution(HR)images is increased,their super-resolution performance drops.In this study,we design a relay residual recurrent neural network to enhance the reconstruction effect of super-resolution reconstructed images,by cascading two deep residual networks.The reconstruction network in charge of coarse reconstruction,and the relay network is further responsible for fine detail reconstruction.Experimental results over SpaceNet satellite image dataset demonstrate that the proposed algorithm is superior to other deep learning approaches,including convolutional neural network super-resolution.2.Satellite imagery super-resolution based on multi-scale residual deep neural network: For existing super-resolution networks,a small receptive field is used for fine reconstruction details.Large-/middle-/small-scale deep residual neural networks are designed simulate different size receptive fields with human visual experience “look in multisacle to see better”.The three different scale networks are designed as different levels of visual tasks to acquire relative global/contextual/local information.Neural networks are then used to fuse different level high-frequency information.In addition,the feature matching based on super-resolution reconstruction is used to verify that the reconstruction results of this algorithm are superior to the frontier satellite image super-resolution algorithm in subjective quality,objective scoring,and recovery feature’s degree.