DeFusion: Aerial Image Matching Based on Fusion of Handcrafted and Deep Features

Machine vision has become a crucial method for drones to perceive their surroundings, and image matching, as a fundamental task in machine vision, has also gained widespread attention. However, due to the complexity of aerial images, traditional matching methods based on handcrafted features lack the ability to extract high-level semantics and unavoidably suffer from low robustness. Although deep learning has potential to improve matching accuracy, it comes with the high cost of requiring specific samples and computing resources, making it infeasible for many scenarios. To fully leverage the strengths of both approaches, we introduce DeFusion, a novel image matching scheme with a fine-grained decision-level fusion algorithm that effectively combines handcrafted and deep features. We train generic features on public datasets, enabling us to handle unseen scenarios. We use RootSIFT as prior knowledge to guide the extraction of deep features, significantly reducing computational overhead. We also carefully design preprocessing steps by incorporating drone attitude information. Eventually, as evidenced by our experimental results, the proposed scheme achieves an overall 2.5–6x more correct matches with improved robustness when compared to existing methods.