A Novel YOLOv5s-Based Tiny Object Detection Model for UAV

The diversity of target scales and the complex background interference present in UAV imagery pose significant challenges for UAV target detection. In response, this paper proposes a novel small object detection model built upon the YOLOv5s. We introduce a new layer, Head-P2, dedicated to detecting small targets within the YOLOv5s framework to enhance precision. Additionally, we incorporate a decoupled detection head to refine the model's target positioning capability. Further augmenting the model's capabilities, the SimAM attention mechanism is integrated into YOLOv5s' feature extraction and fusion network, amplifying feature attention on targets and improving overall feature extraction performance. To address the limitations of the CIOU loss function, particularly its inability to adapt to varying detection targets and challenges with learning difficult samples, we adopt the inner-CIOU loss function. This adaptation enhances both precision and training convergence speed of the YOLOv5s. The modified model is rigorously evaluated on the VisDrone2019 dataset, with ablation experiments confirming its efficacy. In comparative experiments, our enhanced model exhibits notable improvements over YOLOv5s, with a 6.7% increase in [email protected] and a 3.8% increase in [email protected]:.95 values. Moreover, it outperforms YOLOv8s and other models, demonstrating its robustness in small target detection tasks.