Lightweight Small‐Object Detection in Complex UAV Scenarios: A Hierarchical Multi‐Scale Adaptive Approach

ABSTRACT

In the field of UAV small‐object detection, issues such as low resolution, complex background occlusion, and drastic variations in object scale are particularly prominent, becoming core challenges for the application of computer vision in this area. To address these challenges, this paper proposes a lightweight small‐object detection algorithm, HMSA‐Net, suitable for complex UAV scenarios. First, a collaborative mechanism combining hierarchical feature differentiation enhancement with multi‐scale feature complementarity is designed. This mechanism can simultaneously strengthen the advantages of features at different levels. Second, an adaptive dynamic activation strategy is introduced. Under the premise of strictly controlling computational complexity, this strategy can significantly improve the model's ability to capture key features of small objects. Next, the P2 shallow feature map output of the backbone network is reorganized at a sub‐pixel level; at the same time, a frequency‐space collaborative attention mechanism is integrated to achieve early retention and enhancement of fine‐grained details of small objects, further enabling the refined fusion of enhanced shallow features with upsampled deep features across layers. Finally, a small object‐aware statistical enhancement strategy is proposed, which adaptively adjusts the distribution of small‐object features, simultaneously optimizing the model's localization precision. The experimental results on the VisDrone2019 validation set showed that compared with the baseline YOLO11n model, the HMSA‐Net model's precision increased by 5.6%, recall increased by 4.0%, and mAP50 increased by 5.4%. On the test set, the corresponding metric improvements were 3.6%, 3.4%, and 3.7%, respectively. Further evaluation on TinyPerson, UAVDT, and NWPU‐VHR‐10 datasets confirms the superior generalization ability and performance of the algorithm.