Abstract
This study proposes an improved Residual Asymptotic Feature Pyramid Network (R-AFPN) to address challenges in small target detection from the Unmanned Aerial Vehicle (UAV) perspectives, such as scale imbalance, feature extraction difficulty, occlusion, and computational constraints. The R-AFPN integrates three key modules: Residual Asymptotic Feature Fusion (RAFF) for adaptive spatial fusion and cross-scale linking, Shallow Information Extraction (SIE) for capturing detailed shallow features, and Hierarchical Feature Fusion (HFF) for bottom-up incremental fusion to enhance deep feature details. Experimental results demonstrate that R-AFPN-L achieves 50.7% AP(50) on the TinyPerson dataset and 48.9% mAP(50) on the VisDrone2019 dataset, outperforming the baseline by 3% and 1.2%, respectively, while reducing parameters by 15.1%. This approach offers a lightweight, efficient solution for small target detection in UAV applications.