SGCM: Semantic and Geometric Consistency for Robust Aerial Image Matching

Aerial image matching remains an ongoing challenge due to the significant geometric distortions and nonlinear radiation distortion. Most current approaches primarily focus on low-level image features, while overlooking the potential of high-level semantic information to guide the matching process, limitting the robustness and accuracy of these methods in complex scenarios. In this paper, we propose a novel coarse-to-fine aerial image matching framework called SGCM, which considers both semantic and geometric consistency in feature representation. Concretely, in the coarse matching stage, the SAM foundation model is leveraged to efficiently extract salient structural regions, providing valuable semantic guidance for subsequently matching procedure. To achieve accurate region matching, we construct the region description graphs that incorporating both semantic categories and geometric attributes information. A graph neural network (GNN) is then employed for feature aggregation and node descriptor updates, leading to more robust region representations. Coarse matching is performed by computing a similarity matrix between nodes descriptors, which helps effectively mitigate rotational and scaling differences between input images. Consequently, a fine matching strategy combining intra-region point matching and global matching is developed, which significantly improves the robustness of image matching under geometric variations. The performance of our method has been thoroughly evaluated on three challenging aerial image datasets and compared with five state-of-the-art matching methods. Extensive experimental results demonstrate the substantial advantages of SGCM in terms of matching accuracy and efficiency.