Abstract: Scene matching is a key technology for solving the problem of autonomous positioning of aircraft in satellite navigation denied environments. It is of great value for improving the positioning capability of aircraft in visually rich regions and supporting their reliable application in highly dynamic environments such as near space. Existing deep learning-based methods struggle to effectively distinguish stable intrinsic structures from transient noise in images, resulting in insufficient generalization ability when faced with complex domain changes, e.g., drastic changes in viewpoint, season, and modality. Furthermore, they lack explicit physical prior guidance to ensure robustness. As a consequence, this paper proposes a low-rank feature enhancement-based scene matching method. By embedding low-rank priors into a deep neural network, an end-to-end Low-rank Feature Enhancement Network (LFE-Net) framework is constructed. The Schatten-p norm loss implicit constraint model focuses on stable scene structures, and multi-task learning is combined to improve generalization performance. Experiments show that this method achieves higher average localization accuracy on aircraft scene matching datasets and exhibits strong robustness to complex domain changes.

Key words: scene matching, visual geo-location, end-to-end learning, low-rank feature enhancement