Air-YOLO: A UAV-based Target Detection Model for Aerial Photography

Object detection in unmanned aerial vehicle (UAV) imagery has found widespread applications in disaster response, target tracking, and industrial inspection. While UAVs offer advantages such as low cost and flexible deployment, their captured images often contain a high proportion of small objects, frequent occlusions, and complex backgrounds, posing significant challenges to detection algorithms. To address these issues, we propose Air-YOLO, a lightweight and robust detection framework based on the YOLO architecture, tailored for UAV scenarios. The proposed method incorporates a small-object feature enhancement module (SFEM) to strengthen the interaction between shallow and mid-level features, thereby improving small-object representation. In the backbone, deformable convolution-based residual blocks (C2f_DCN) and channel prior convolutional attention (CPCA) are introduced to enhance geometric adaptability and preserve long-range contextual information. Furthermore, a partial convolution enhanced head (PCE-Head) is designed to reduce computational redundancy while improving detection accuracy. Extensive experiments on VisDrone and CODrone datasets demonstrate that Air-YOLO achieves higher detection accuracy than state-of-the-art baselines, while reducing parameter count and GFLOPs. These results highlight the model’s efficiency and robustness, showcasing its potential for real-time UAV-based applications.