关键词: 线特征匹配, 跨视角影像, 神经网络, 灰度信息, 特征融合

Abstract: The geometric deformation of cross-view images is large, and the unilateral neighborhoods of corresponding lines are prone to significant differences, making it difficult for traditional line matching algorithms to obtain reliable line pairs. To solve this problem, a line matching algorithm for cross-view images combining neural network learning with grayscale information is proposed. Firstly, the pixel-level directional gradient histogram features of the images are obtained and a neural network is applied to fuse these features and the image greyscale information to form a feature description grid. Then, the discrete points are extracted from the line, and the one-sided descriptions of the points are calculated according to the information of the discrete points' unilateral features description grid. The one-sided descriptors of the points are aggregated by the deep learning method to form the unilateral abstract expression of the line. Finally, the known corresponding points are used to constrain the matching region for group matching, and the matching results are verified by comparing the topological consistency of the line pairs to obtain the final matching line pairs. A number of representative groups of public cross-view images from different sources are selected for line matching experiments and they are compared with mainstream line matching algorithms for analysis. The results show that the proposed algorithm is able to obtain uniformly distributed line pairs with high correct rates for cross-view images from different sources with obvious differences in content, and realize robust line matching of cross-view images from different sources.

Key words: line matching, cross-view images, neural network, grayscale information, feature fusion