Remote Sensing Image Matching Based On Deep Learning

Remote sensing imaging technology is widely used in the fields of target location,geological survey,and detection of environmental changes on the earth's surface.Image matching is one of the research hotspots in remote sensing imaging technology.Due to factors such as uneven illumination distribution,hardware equipment displacement and obstacle occlusion during the generation process of remote sensing images,the accuracy and speed of image matching will be affected.Convolutional neural networks in deep learning have powerful capabilities in feature extraction and learning,and robust to noise,deformation,and illumination changes,they can improve the accuracy of image matching.Therefore,the use of deep learning methods to match remote sensing images is of great significance for research and practical application.Firstly,the research background and development status of remote sensing image matching are introduced.The basic concepts,common methods,matching strategies and performance evaluation of image matching are expounded.The basic concepts of deep learning and convolutional neural networks are outlined.Moreover,the main training strategies,and the commonly used data enhancement methods are summarized.Secondly,the methods of remote sensing image matching based on deep learning are studied.The specific content is as follows:(1)The remote sensing images matching network based on the RF-Net(Receptive Fields Network)and SIFT(Scale Invariant Feature Transform)algorithm is studied.This method integrates the traditional matching algorithm into the deep learning network framework,uses the feature maps extracted by the convolutional neural network to construct the scale space and the Gaussian difference space,and obtains the information of keypoints,and finally generates a descriptor for the matching of remote sensing images.Experimental results show that the matching method based on deep learning is superior to the traditional SIFT method;the matching method based on RF-Net is superior to the matching method based on RF-Net and SIFT.(2)A matching network based on RF-Net and DSS(Deeply Supervised Salient Object Detection with Short Connection)is proposed.In the keypoints detection backbone network,the short-connection structure is introduced to transfer the high-level features to the shallow side output,which can generate response maps with more information,for extracting keypoints information.and introduce a descriptor network similar to L2-Net to obtain 128-dimensional feature vectors to describe keypoints.Experimental results show that the matching performance of this method is slightly better than the method based on RF-Net,and the keypoints detection effect is not as good as the latter.(3)A dual-channel matching network based on RF-Net and PSPNet(Pyramid Scene Parsing Network)is proposed.This method introduces PSPNet for the problems of shallow network and lack of high-level semantic information in RF-Net,the dual-channel network framework was built to aggregate the context information on different regions,and merge the shallow network with increasing receptive fields,generates descriptor for remote sensing image matching.Experimental results show that the matching accuracy of this method is slightly weaker than RF-Net,but it has a greater improvement in the performance of anti-affine changes.(4)A dual-channel matching network based on RF-Net and Res Net(Residual Network)is proposed.The two channels of this method are composed of RF-Net and ResNet,respectively generates shallow features with increasing receptive fields,and high-level features with stronger representation capabilities.These two features are combined to obtain a feature map containing rich information on the premise of ensuring the size of the receptive fields.The experimental results show that the method has better performance in the detection and matching accuracy of keypoints in remote sensing images.Finally,Summarize the content of the full text and look forward to the future research directions.