Super-resolution Reconstruction Of Remote Sensing Images Based On Reference Feature Transfer

Remote sensing image super-resolution reconstruction is an image enhancement method based on computer vision and deep learning techniques,which aims to generate high-resolution images from low-resolution remote sensing images.The main purpose of this technology is to improve image quality,increase the detail texture and clarity of remote sensing images,and better support related research such as map making,earth science,and natural resource management.However,due to the limitations of image transmission and device imaging conditions,high-resolution remote sensing images of specific regions and time ranges are usually difficult to obtain,limiting the application of remote sensing images in specific fields such as geological exploration and urban planning.Existing remote sensing image super-resolution methods can learn features from a large number of remote sensing images,extract multi-scale texture information,and generate high-resolution remote sensing images,achieving good results in various scenarios.However,current reconstruction methods still have problems such as unclear detail textures and difficulty in reconstructing missing information.Therefore,some studies have introduced high-resolution reference features of the same region’s remote sensing images to provide reference information for the super-resolution reconstruction process,which can restore the lost texture details in low-resolution images.However,due to the color and brightness differences between the reference image and the original image,this method still results in color and brightness imbalances and pseudo-artifacts in some areas,limiting the promotion and use of reference-based image super-resolution methods in the field of remote sensing image reconstruction.In addition,a remote sensing image covers all the land cover information of that region,and the image scales of different objects in the scene and the types of land cover are complex and diverse.Most of the super-resolution reconstruction methods for natural images have not considered these characteristics of remote sensing images and lack general applicability and effectiveness in solving remote sensing super-resolution reconstruction problems.To address the above problems,this paper proposes a remote sensing image superresolution reconstruction method based on reference image alignment and feature adaptive transfer,using remote sensing images collected from different satellites as reference information,guiding deformable convolution with optical flow for feature alignment,and using spatial adaptive methods and convolutional block attention modules for feature transfer.While improving the resolution of remote sensing images,this method achieves consistency between the color and brightness distribution of reference image features and low-resolution image features,avoiding the influence of irrelevant reference information on image reconstruction.According to the differences in features of remote sensing images in different scenes,this paper conducts targeted training on images in specific scenes,achieving better training results than traditional single-scene training.The main research content and contributions include the following three aspects:(1)To eliminate the interference caused by the differences in ground objects between two remote sensing images and the deviation of the shooting angle on super-resolution reconstruction,this paper proposes a feature alignment method for remote sensing images:using optical flow to guide deformable convolution to align image features.This method can effectively detect the deviation of ground object positions in remote sensing images from the same region but from different sources,and can be annotated through visualization,playing a role in maintaining image clarity and eliminating texture interference after reconstruction.(2)Due to the differences in acquisition equipment and time,there are obvious differences in color and brightness between remote sensing images,which results in a large difference between the reconstructed image and the original image.This paper introduces a spatially adaptive de-regularization method and an improved convolutional block attention module to achieve the consistency of color and brightness distribution in the reconstructed remote sensing image features and avoid the adverse effects of irrelevant reference information on remote sensing image super-resolution reconstruction.(3)This paper designs an image alignment and feature transfer reconstruction network based on scene classification,which integrates the optical flow-guided deformable alignment module,reference feature transfer module,and scene-adaptive classification reconstruction method into the same super-resolution reconstruction network.The scene-adaptive reconstruction method is used to fine-tune the pre-trained network using remote sensing images from different scenes after obtaining the pre-trained basic network.This method can improve the reconstruction effect in multiple scenes,especially for remote sensing image scenes with complex texture structures,with more significant improvements in peak signal-to-noise ratio and structural similarity evaluation indicators.The experimental results show that the proposed method can effectively improve the color and texture of the reconstructed image and eliminate the adverse effects of irrelevant reference information on remote sensing image super-resolution reconstruction in the reference-based remote sensing image super-resolution reconstruction dataset.In situations where there are differences in color and brightness and pixel-level deviation between the reference remote sensing image and the remote sensing image to be reconstructed,the interference caused by the reference image difference can be effectively eliminated,achieving better visual performance.The proposed method can effectively solve the problems existing in remote sensing image super-resolution reconstruction.