Aerial Image Matching Method Based On HSI Hash Learning

For aerial image high-dimensional,high-resolution,large size,etc.,a Hash-based image matching algorithm(HSI-Hashing)based on HSI color space is proposed in this paper to achieve fast matching of high-resolution aerial images.Firstly,from the perspective of human eye vision,the original image is converted to HSI component image,and then the feature point is detected and a multi-descript descriptor is formed.Then,the hash objective function is constructed with inter-class and intra-class hash information,and the hash functions can be learned.Finally,the high-dimensional feature description vector is mapped to the binary hash code through the hash function,and the fast matching is realized in the Hamming space.The main work is as follows:(1)Traditional SURF,SIFT and other algorithms are mostly based on gray level information for feature detection,and their neglect of image color or edge information easily leads to missed matching.Therefore,this paper proposes to transform the original RGB image into HSI image,make full use of its hue,saturation and brightness information for feature selection and description.In order to deal with the abundant edge information of aerial images,an margin non-maximum suppression M-NMS method is proposed,which greatly improves the accuracy of feature point selection.(2)In order to improve the ability of feature description,a local multi-feature descriptor based on HSI color space is proposed.The average direction feature is extracted on the H component,the gradient feature is extracted on the S component,and the average intensity feature is extracted on the I component to form a highly robust feature description vector.Descriptor performance is verified by Precision-Recall curves and matching score experiments.(3)In order to achieve fast matching of high-dimensional feature descriptors,binary coding mapping is performed by a hash function.Firstly,to ensure that the sample points of different categories are more independent,the stronger the correlation of the same category,that is,the inter-class hash is greater and the intra-class hash is smaller.Based on this,a hash learning objective function is constructed.Secondly,in order to accelerate the algorithm parameter determination,the projection matrix and the offset threshold are determined by thegeneral relaxation,Lagrangian multiplier and similarity information.(4)In order to ensure the similarity between European space and Hamming space,this paper proposes a minimum distance error loss function based on Hamming distance and Euclidean distance,and sets the Hamming matching threshold for each component image.At last,the effectiveness of the aerial image matching algorithm proposed in this paper is verified by the standard database and the real aerial image of the UAV.The three-dimensional reconstruction of the aerial target and the general target image is also used to verify the versatility and robustness of the algorithm.