Remote Sensing Image Fusion Based On Attention

Remote sensing satellites provide people with a representation of the Earth’s surface.Using remote sensing satellites to inspect key areas regularly and analyze anomalies;To carry out a survey of unfamiliar areas and obtain topographic features.Regular early warning,action guidance and support,and geographic information support will be implemented to provide guarantees for security and stability.Limited by the need to balance the spatial resolution and spectral resolution of images acquired by a single sensor,many satellites are equipped with two kinds of optical sensors to provide two complementary data,such as multi-spectral images with low spatial resolution but rich spectral information and panchromatic images with high spatial resolution and most serious spectral information loss.The proposed pan-sharpening technique solves how to get the image with high spatial resolution and high spectral resolution.Pan-sharpening refers to the use of high spatial resolution panchromatic image to sharpen the multi-spectral image of low spatial resolution,so as to obtain the multi-spectral image with both high spatial resolution and high spectral resolution.It is the pre-processing of remote sensing image target detection and remote sensing segmentation and other advanced processing.The main contents of this paper are as follows:(1)A pan-sharpening framework based on Transformer regression network is proposed to integrate multi-spectral image and panchromatic image.This network consists of backbone network and double regression network.Backbone network for feature extraction,feature fusion and image reconstruction;Double regression network can optimize the network and improve the performance of the network by calculating the loss of the shallow feature image and the result image after downsampling in the process of image reconstruction.Swin-Transformer module is added to the feature extraction part instead of subsampling to avoid the loss of pixel-level internal structural features and increase the robustness of the feature model,thus reducing the complexity of the network.The image reconstruction part combines the convolution block attention module and the effective channel attention module to focus on the spectral information of the multi-spectral image and the spatial information of the panchromatic image,preserves the spectral information and spatial details to a greater extent,and makes the pan-sharped image more close to the real image.(2)Proposed a remote sensing image based on attention and iteration in stages networks depth information fusion method,the iteration of each network as a phased sharpening process,and in each subnet for feature extraction and image reconstruction operation.Using the advantage of the iterative network to the depth features of the cross-stage,the multi-spectral image and panchromatic image refinement features are extracted hierarchically for image reconstruction.In the feature extraction stage,we use the large kernel attention module and the cascaded asymmetric coupled representation module as the feature extraction backbone.The large core attention module has channel and spatial adaptability,as well as strong remote dependence establishment ability,which makes feature extraction more complete.The asymmetric coupling representation module outputs the refined spectral and spatial features to the next stage by learning the mixed correlation of multispectral and panchromatic images.In the stage of image reconstruction,spectral and spatial features are fused using the attention fusion module.The input spectral and spatial features are effectively used to reduce information loss and retain important information.Finally,features are extracted more completely and high quality reconstructed images are obtained to obtain ideal fusion images.