PHAF-YOLO: lightweight progressive multi-kernel enhancement and hierarchical adaptive fusion for real-time aerial small object detection

Abstract

Aerial small object detection is confronted with critical challenges including extreme scale variation, dense target distribution, heavy background interference and strict real-time deployment constraints on unmanned aerial vehicle platforms. Traditional real-time detectors fail to balance detection accuracy and computational efficiency in such scenarios, suffering from severe cross-layer feature misalignment and small object feature attenuation caused by the mismatch between deep semantic features and shallow spatial features. To address these issues, this paper proposes a lightweight real-time detection framework for aerial dense small objects, named PHAF-YOLO. Three targeted optimizations are designed: the Progressive Multi-Kernel Enhancement Unit is embedded in the backbone to expand the effective receptive field with low overhead and alleviate deep feature attenuation via progressive multi-kernel convolution; the Hierarchical Dual-stage Adaptive Fusion Block is applied in the neck to dynamically screen multi-scale features and suppress fusion redundancy through dual-stage adaptive fusion, improving cross-scale information utilization for dense small objects; the Target-Centric Soft Regression Loss combines object-centered geometric constraints with soft target assignment to mitigate gradient instability from ambiguous positive-negative sample boundaries and boost high-IoU localization accuracy. Extensive experiments on VisDrone2019, UAVDT2018 and AI-TOD2021 demonstrate that PHAF-YOLO achieves a superior accuracy-efficiency trade-off across all model scales and outperforms mainstream real-time detectors. Ablation and validation experiments further verify the independent effectiveness and synergistic effect of all core modules. The source code of this work is publicly available at: https://github.com/csy001x/PHAF-YOLO.