Structural Priming With High-Frequency Gradient for Mamba in UAV Object Detection

Low-altitude unmanned aerial vehicles (UAVs) are increasingly employed in surveillance and monitoring tasks such as perimeter inspection, urban patrol, and emergency response. Object detection in these scenarios is highly challenging due to the small size of targets, frequent occlusions, cluttered backgrounds, and blurred boundaries from aerial viewpoints. Existing detectors primarily emphasize semantic abstraction, while overlooking structural cues such as edges and contours, which are often critical for precise localization. To address this gap, we propose SPHMamba (Structural Priming with High Frequency Gradient for Mamba), a boundary-aware detection framework that explicitly integrates semantic and structural information. A dual-stream representation module first extracts complementary features through a Semantic Stream and a Structural Stream, yielding both detection predictions and modulation matrices. These semantic and high-frequency structural priors are subsequently injected into a Structure-Primed Mamba Block (SPMB), where the selective scan mechanism adaptively propa gates boundary-sensitive information and suppresses background noise, producing a High-Probability Correction Matrix (HPCM). The refined representation is then incorporated by a Boundary Refinement Module (BRM), optimized with a Structural Alignment Loss, to enhance localization precision and achieve tighter alignment with object contours. Extensive experiments on challenging UAV benchmarks demonstrate that SPHMamba consistently outperforms state-of-the-art detectors in both detection accuracy and boundary localization, providing an efficient and generalizable solution for intelligent UAV-based monitoring in complex environments.