Research On Super-resolution Reconstruction Of Optical Satellite Images

In recent years,with the progress of satellite image processing technology,optical remote sensing images have been more and more widely used in real scenes,and high-resolution remote sensing images play an important role in many fields such as urban planning,semantic annotation and target detection.However,due to the limitations of current remote sensing imaging technology and the influence of external environment,the resolution of remote sensing images is often difficult to meet the application requirements.In order to obtain high-resolution remote sensing images,image super-resolution(SR)methods are gradually applied to the recovery and reconstruction of remote sensing images.Most existing studies have used single-image super resolution(SISR)methods to reconstruct high-resolution(HR)images.However,due to the lack of information in low-resolution(LR)images and the unsuitability of the SISR method,it is difficult to reconstruct the fine textures of HR images at large magnifications(e.g.,four times).To address the above problems,this thesis conducts a related study based on the reference image super-resolution(Ref SR)method,which uses the fine textures in high-resolution reference images to compensate for the lack of details in the original LR images to assist in improving the performance of the SR method.In addition,most accessible public remote sensing datasets cannot maintain both long time coverage and high spatial resolution.For this reason,two remote sensing image datasets are constructed in this thesis to support the smooth implementation of Ref SR work.The main elements are as follows:(1)Construction of remote sensing image datasets based on reference images:Aiming at the problem of missing remote sensing image datasets based on reference images,this thesis adopts the longitude and latitude matching and slicing methods to construct a set of HR-LR remote sensing grayscale image datasets with the publicly available Google online map high-resolution image as HR image and the publicly available Argcis online map image as LR image;with the publicly available Google online map high-resolution image as HR image,the A set of HR-LR remote sensing color image dataset is constructed with the public Google online map high-resolution image as the LR image,and with the high-fraction-1 satellite image as the LR image.It provides data supplement and support for the research application of Ref SR method in the field of remote sensing.(2)Proposed algorithms for super-resolution of remote sensing images based on residual dense hybrid attention networks:The SR model proposed in this method uses the super-resolution of neural texture transfer(SRNTT)as the backbone network,and based on this structure,we propose a dense hybrid attention block(DHAB)as the building block of the residual dense hybrid attention network(R-DHAN).The DHAB incorporates the input of the current block and its internal features.While making full use of the feature information,the interdependencies between different channels and different spatial dimensions are used for modeling,and a strong representation capability is obtained.In addition,a hybrid channel spatial attention mechanism is introduced to focus on important and useful regions for better reconstruction of the final image,and experiments show that the proposed R-DHAN method performs well in terms of quantitative evaluation and visual quality compared with SRNTT and some classical SR techniques.(3)Proposed algorithms for super-resolution of remote sensing images based on improved texture transformation networks:To further capture and exploit the high-frequency information present in LR space,an improved texture transformation network(ITTN)is proposed to reduce the redundant features in the reference image and thus retain more useful feature information.In addition,a new backbone network,named the Remaining Attention Connection Network for Channels(RA-CAN),is proposed to obtain powerful network characterisation capabilities with the help of the channel attention mechanism and attention connection patterns,which can extract rich low-frequency components and valuable high-frequency texture details from the LR space.The experiments show that the proposed ITTN method achieves good performance in both quantitative and qualitative measurements.