Research On High Resolution Aerial Image Matching Based On Deep Hashing

Aerial images have features of high resolution,complex background,and usually contain rich texture and edge information.However,most algorithms used in matching of aerial images adopt the shallow hand-crafted features(e.g.,SIFT,SURF),which may suffer from poor matching speed and are not well represented in the literature.Aiming at the problems that the traditional feature descriptors encountered in high-resolution aerial image matching,and achieving its effective and rapid matching,this paper has carried out in-depth research and proposed a Local Deep Hashing Matching(LDHM)method for matching of aerial images.Firstly,according to the course overlap rate of aerial images,the optimal local matching region is extracted;Secondly,uniformly distributed feature points are detected in the local matching region,and the neighboring image patches of feature points are extracted;Thirdly,image patches are imported to the local deep hash network,thus obtaining the representation of a binary hash code;Finally,similarity of the image patches is compared in the Hamming space to complete the matching of aerial images.Through comparison and analysis,the proposed algorithm improves the matching efficiency and meets the real-time requirements under the condition of ensuring the matching accuracy.In this paper,the main content and innovation points are as follows:(1)According to the course overlap rate of aerial images and the normalized cross-correlation algorithm,a matching method based on aerial prediction region is proposed,which constructing the local matching region to reduce the computational complexity and improve the matching efficiency.(2)A uniformly distributed feature point extraction algorithm based on local matching region is proposed.The local matching region is meshed and the candidate meshes are determined according to the information entropy to obtain uniformly distributed feature points.(3)Based on the Triplet network,a local deep hash matching framework for aerial image matching is constructed:learning the deep features of image patches through convolutional neural network,and constructing the hash layer to obtain hash codes with independence and similarity.(4)The absolute distance constraint is increased in the traditional Triplet loss to overcome the shortcomings that the positive and negative samples in the relative distance constraint of Triplet loss cannot optimize the network when the distance is greater than the margin parameter.Besides,considering the quantization loss in the objective function,to obtain hash codes with better discrimination and characterization.