Subtle-YOLOv8: a detection algorithm for tiny and complex targets in UAV aerial imagery

Unmanned Aerial Vehicle (UAV) imagery for small target detection plays a crucial role in traffic safety, military defense, and agricultural production. Despite rapid advancements in target detection algorithms, tiny targets like pedestrians, people, and bicycles still encounter significant challenges in practical applications, including occlusions, low resolution, and difficulties in capture and segmentation. These challenges require detectors to be highly adaptive and capable of precisely distinguishing between targets and dynamic backgrounds. To address these issues, we use YOLOv8 as the baseline model and proposes a new detection network named Subtle-YOLOv8. Initially, dynamic snake convolution (DSConv) is incorporated into the backbone network to enhance the perception of subtle information and feature extraction efficiency. Secondly, an attention mechanism called Efficient Multi-scale Attention Module (EMA) is introduced to optimize the neck network to improve the transfer of key features. Finally, we designed a tiny object detection head and replace the original loss function with Wise-IoU, focusing the model more on samples of ordinary quality and further enhancing the detection capabilities for tiny targets. Experimental results show that our model achieves a 6.2% improvement in average detection precision over the baseline with a slight increase in parameters. It particularly excels in handling complex tiny targets such as pedestrians and people, with detection precision improvements of 14% and 12%, respectively. The code will be soon released at https://github.com/WilliamXSS/SubtleYOLO