Aerial Small Object Detection with Boundary-Aware and Semantic-Guided Feature Pyramid

Abstract

Aerial small object detection in UAV imagery faces multiple challenges, including highly imbalanced scale distribution, low object resolution, complex background textures, and constrained computational resources. To address these issues, this paper proposes Aerial-YOLO, an aerial object detection algorithm that integrates boundary-aware enhancement and a semantic-guided feature pyramid. First, a boundary-aware and scale-adaptive collaborative enhancement module is constructed in the backbone network. By introducing directional structural reinforcement and adaptive multi-scale convolution mechanisms, the proposed module strengthens object contours and scale-discriminative features, thereby improving localization stability in dense scenes. Second, a lightweight semantic-guided feature pyramid network is designed, whose core component is an Adaptive Cross-scale Injection Gate module. Through unidirectional semantic guidance and sparse fusion strategies, high-level semantic information selectively enhances low-level fine-grained details, improving multi-scale representation capability while reducing computational overhead. In addition, a structural detail reconstruction module is embedded between the backbone and the neck, where directional enhancement and frequency-domain reweighting mechanisms are employed to amplify high-frequency responses of small objects and alleviate interference from complex backgrounds. Experimental results on the VisDrone2019 dataset demonstrate that Aerial-YOLO significantly improves small object detection performance while maintaining real-time inference capability. Further evaluations on the TinyPerson dataset verify the strong generalization ability and robustness of the proposed method.