Research On Aerial Image Matching Method Based On Deep Learning

The image matching method is a method to solve the corresponding relationship between the common elements between the search graph and the template graph,and has important application value in image registration,target tracking and other issues.This paper studies the matching method of aerial images,which can be widely used in military research,academic research,life medical and other fields.Aiming at the difficulties that the aerial image matching problem is more sensitive to illumination and viewing angle changes than the traditional image matching problem,this paper designs two methods with different technical routes based on deep learning technology,which are based on the image feature matching method MS-FPM(MultiStage Model Based on Feature Point Matching)and IM-TL(Image Matching Based on Target Location),the IM-TL method is the first aerial image matching method using image target location technology.The experimental results show that the MS-FPM and IM-TL methods can improve the matching accuracy of aerial images under complex illumination and viewing angle changes,and have good practical application value.This paper completes the work as follows:(1)This paper constructs the aerial positioning data set Data＿IM.The data set is constructed based on the aerial photography data set GL3 D.In this paper,the size of the image in the data set is unified to 640x480 as the search image,and the minimum circumscribed rectangle of the target is intercepted in the search image and the size is unified to 127x127 as the template image,followed by random affine transformation,to obtain 5000 pairs of aerial images containing 378 different scenes,and the positional relationship annotation information between the image pairs.Data＿IM is an aerial positioning dataset with rich scene categories and large data scale at this stage.(2)This paper proposes a multi-stage aerial image matching method MS-FPM based on feature point matching.MS-FPM includes three steps: feature point extraction,feature point vector description,and feature point matching to complete image matching.In the feature point extraction stage,the DOG operator is used to extract the key points of the image pair.In the feature point vector description stage,this paper proposes the EL2-Net(Effective-L2-Net)model as the descriptor extraction model in the MS-FPM method.Multiple Conv Blocks are added to the EL2-Net to solve the problem of insufficient shallow image information.problem,dig deep into image information,and get more comprehensive feature descriptors.Aiming at the unbalanced ratio of positive and negative samples in the input samples,this paper uses contrastive loss as the loss function of EL2-Net to improve the network’s degree of discrimination of samples.In the feature point matching stage,the descriptor in the template graph finds the nearest neighbor descriptor in the search graph through the BBF algorithm and calculates whether its nearest Euclidean distance is less than the threshold to obtain the matching result.The experimental results show that on the aerial data set,the matching accuracy of MS-FPM reaches 78.54%,which is 1.65%and 3.02% higher than the deep learning methods R2D2 and D2-Net,respectively,and the average matching time of MS-FPM is 0.96 seconds.Compared with the deep learning methods R2D2 and D2-Net,it saves 0.27 seconds and 0.42 seconds,respectively.(3)This paper proposes an image matching model IM-TL based on object localization.The model IM-TL uses the improved Alex Net module to integrate the ECANet attention mechanism in the feature extraction stage to improve the feature extraction ability of the model.IM-TL inputs the classification branch and regression branch of the network obtained by the feature extraction module into the cross-correlation module,and then obtains the predicted position of the matching target and the corresponding positive sample confidence through the network,and finally obtains the final target through the argmax operation.Image matching position.In order to improve the accuracy of target position matching,this paper uses DIOU as the loss function in the classification branch.The experimental results show that on the positioning dataset Data＿IM,when the matching threshold is set to 6 pixels,the matching accuracy of the IM-TL model is 0.94% and 0.56% higher than that of the R2D2 model and MS-FPM model,respectively.This paper compares the proposed MS-FPM method and IM-TL model with the existing model R2D2,which performs the best image matching task on the GL3 D dataset,on the public image matching dataset ROxford.The experimental results show that the image matching accuracy of MS-FPM method,IM-TL model and R2D2 model are 63.11%,63.83%,and 64.41%,respectively,which verifies the generalization effect of the method proposed in this paper on image matching tasks.