An effective detection algorithm for small UAV based on lightweight You-Only-Look-Once (YOLOv4-L) approach

The control and monitoring of small Unmanned Aerial Vehicles (UAV) plays a crucial role in national defense and security. However, due to their compact size and high mobility, the detection of small UAV across diverse scenarios remains a significant challenge. To address this issue, this study proposes an improved detection algorithm tailored for small UAV. The model is initially trained on a virtual dataset, and the learned parameters are transferred to real-world data through a transfer learning framework. To optimize anchor box generation, clustering analysis is performed on bounding box dimensions, resulting in anchor boxes with appropriate scales and aspect ratios. Furthermore, the Ghost module is introduced to replace conventional convolutions in CSPDarknet53, enhancing feature extraction efficiency. An Efficient Channel Attention (ECA) mechanism is also incorporated to strengthen output feature representations and improve the capture of texture details critical for small target detection. Through experiments, the proposed algorithm can achieve the mAP0.5 of 82.2 %. Experimental results demonstrate the effectiveness of the proposed small UAV detection method.