Fast Aerial Image Matching Based On Deep Learning

Image matching plays an important role in UAV visual navigation.The position and speed information of the aircraft can be obtained by matching the aerial image with the electronic map,or by matching the two adjacent aerial images.The traditional image matching algorithms are mainly based on hand-craft features with good generalization,but the matching performance may become poor in some tasks.Deep learning is developing rapidly at present,but just rising on image matching,and there is a large space for improvement in accuracy and speed.Therefore,this thesis mainly studies image matching algorithms based on deep convolution neural network,in order to improve the matching accuracy and speed.The main work is as follows:(1)A method of producing aerial image matching dataset is proposed to improve the matching accuracy of the Hardnet network.Hardnet is a deep convolution neural network to extract 128-dimension feature descriptors,but the UBC Phototour dataset used in training is different from aerial images,which affects the actual matching performance.Therefore,the difference between aerial image data and UBC Phototour dataset is analyzed.Then,based on SIFT algorithm,feature extraction and matching of aerial image data are carried out.Then,based on the idea of twin network,50000 pairs of positive and negative samples are generated,and the aerial image dataset is generated in the format of UBC Phototour dataset.Finally,the dataset generated in this thesis is used to train the Hardnet network.Experimental results show that the proposed image matching method based on Hardnet imporves 1.7% accuracy on average and more stable than the traditional SIFT algorithm.(2)A CENet method is proposed to extract regions of interest(ROI)to improve the speed of Hardnet image matching.At present,the detection results of commonly used ROI extraction algorithms are not accurate enough.Therefore,the idea of ensemble learning is incorporated into the traditional deep convolution neural network.By using multiple deep convolution modules to extract and integrate features,more accurate detection results are obtained.In order to enhance the difference of features extracted by different convolution modules,new training process and loss function are used for training.CENet is constructed with the VGG16 network as the backbone network.The experimental results show that CENet can effectively detect different features and detect more accurate ROI,Compared with the CENet backbone network VGG16 network and the promoted-UNet network,the F value of CENet network is increased by 3.86% and 3.23% respectively,and on MAE,CENet is 14.8% lower than VGG16 network.(3)The CENet and Hardnet networks are combined to realize fast matching of aerial images.Gaussian difference scale pyramid is used to extract feature points,the CENet network is used to extract ROIs in aerial images,and then feature points are selected based on the ROIs.Considering that it takes a long time to annotate the aerial image data manually,this thesis uses LC algorithm to pre-label the aerial images,and then adjusts it manually.Finally,regions around selected feature points are extracted for Hardnet to perform fast matching of aerial image.The experimental results show that the ROI strategy can effectively select appropriate feature points and effectively improve the speed of aerial image matching,1.2 times faster than using hardnet independently and the accuracy of image matching is slightly improved.