Research On Multi-source Remote Sensing Image Matching Algorithm Based On Convolutional Neural Network

Image matching is widely used in automatic driving,satellite navigation,military protection and medical imaging diagnosis and other fields,which is the basis and premise of other image processing tasks,and is also one of the hot spots in the current research of imaging application technology.However,the accuracy of image matching will be affected by various factors such as hardware device displacement,uneven distribution of light intensity,obstacle occlusion and nonlinear radiation during the imaging process.In order to cope with this problem,this paper uses the powerful feature learning and extraction capabilities of convolutional neural networks to extract the deep semantic features of complex images,so as to achieve the purpose of good resistance to image distortion,light contrast changes,nonlinear radiation and other influences,and apply them to the field of remote sensing image matching to improve the performance and efficiency of image matching.The specific research content is as follows:(1)This paper first reviews the research background and development status of remote sensing image matching to understand the scientific background of remote sensing image matching,and on this basis,the basic principles of image matching are expounded,and the common methods of image matching and matching strategies and performance evaluation are briefly introduced.In addition,this paper also summarizes the basic knowledge and network composition structure of convolutional neural networks,the optimization methods of deep learning,and the network training strategies.(2)In view of the shortcomings that traditional algorithms make less use of image deep context information,in order to enhance the match between multi-source remote sensing images,this paper proposes a neural network model based on the matching strategy of "double screening and double constraints",the algorithm improves the RepVGG network to further deepen the feature extraction ability,extracts the image deep semantic feature map through the improved network,and filters the key points of both the principle of priority maximum and the principle of accurate extreme value in the feature map.The strategy of combining the coarse to fine inverse matching constraint and the RANSAC constraint enhances the matching between multi-source images and significantly improves the image matching accuracy and speed.(3)In view of the obvious nonlinear radiation differences between multi-source remote sensing images,the existing deep learning methods ignore the influence of feature relationship learning on feature extraction,and based on this,a fusion feature relationship matching network based on Siamese is proposed.The overall matching network of the algorithm includes a feature extraction module,a fusion feature relationship learning module and a Siamese metric module.The feature extraction module extracts deep features from the input remote sensing image pairs.The fusion feature relationship module introduces the self-attention mechanism fusion feature relationship to capture local context information on the basis of channel dimension feature difference learning and spatial dimension feature product learning.The Siamese metrics module predicts the matching performance of images against the output of the Fusion Feature Relationship module...