MS-AFNet: A multi-scale adaptive fusion network with high-resolution feature enhancement for tiny object detection in UAV imagery

Detecting tiny objects in unmanned aerial vehicle (UAV) imagery remains a formidable challenge due to the diminutive pixel footprint of targets, severe background clutter, and the progressive degradation of spatial fidelity during multi-stage downsampling. This paper proposes MS-AFNet, an advanced deep learning architecture specifically optimized for robust tiny object detection in complex UAV scenarios. To counteract information loss in deep layers, we introduce a Multi-Scale High-Resolution Feature Enhancement (MS_HRFE) module at the P2 level, which effectively captures fine-grained morphological details through parallel multi-scale context branches and a dual-attention mechanism. Furthermore, an Adaptive Spatial Feature Fusion (ASC_FSF) mechanism is integrated to dynamically recalibrate spatial weights during cross-scale feature interactions, thereby suppressing background noise and enhancing adaptability to significant scale variations. Additionally, the RTAS_C3k2 module is designed to construct a parameter-efficient backbone, leveraging spatially sensitive guidance and partial texture enhancement to preserve localization accuracy. Experimental evaluations on the VisDrone2019-DET dataset demonstrate that MS-AFNet significantly outperforms state-of-the-art models. Specifically, evaluated on the rigorous test-dev set, MS-AFNet achieves an mAP 50 of 30.0%, exhibiting an absolute improvement of 7.7% over the YOLO11n baseline. Notably, the proposed model maintains a compact architecture with only 2.4M parameters, achieving an optimal trade-off between detection precision and computational efficiency in resource-constrained environments. This research provides a robust, high-precision detection solution for intelligent urban traffic management and large-scale UAV surveillance.