Research On Feature Matching Method For High Resolution Aerial Image

Aerial images are large in size, high in resolution, and have rich edges and texture information. Image matching algorithm based on traditional floating point feature descriptor has high computational complexity and poor matching speed when dealing with aerial images. In order to realize the fast matching of high resolution aerial images,this paper proposes a Local Multi-feature Hashing (LMFH) method. Firsrly, the prediction area is constructed according to the the heading overlap rate,and the feature points detected in the area are described by multi-feature. Then,the hash functions are learned by a large number of training samples. Finally,the high-dimensional feature description vectors are mapped to compact binary hash codes by the learned hash functions. In the Hamming space, according to the Hamming distance to achieve the fast matching between points. The main research work of the paper are as follows:(1) A matching method based on prediction area is proposed according to the overlap rate of aerial images. And then feature extraction and matching is only operated in the area, greatly reducing the computing time.(2) The construction method of local feature descriptor based on multi-feature is proposed.By combining the characteristics of gradient intensity,grayscale and direction of the neighborhood of feature points, a high robustness feature description vector is constructed.(3) For high-dimensional local feature vectors, a local hashing method is proposed. Firstly,two projection matrices are used to construct hash functions. Secondly,similarity weight is assigned to denoted objective function according to the pairwise labels of local features. Thirdly,alternative iteration method is used to learn the parameters of hash function. At last, quantization strategy is adopted to learn multi-bit binary codes.(4) The local hashing method is introduced into the high resolution aerial image feature matching, and a matching scheme based on Hamming look-up table is proposed, through the Hamming distance to realize fast matching of feature points.