Small-Target Detection Algorithm for Aerial Images Based on YOLO v8

Aiming at the challenges of low resolution, high target density, and background similarity in aerial small-target detection, this study proposes an improved detection algorithm, TINY-YOLOv8. First, to enhance the detection capability for small targets, an additional small-target detection layer is incorporated into the YOLOv8 model, and the backbone network structure is optimized to improve detection accuracy. Second, partial convolutions in the backbone network are replaced with lightweight convolutions, and the C2f module in the neck is substituted with a lightweight VoV-GSCSP module to reduce model parameters and computational complexity. Furthermore, the ECA attention mechanism is integrated into the network to strengthen channel-wise feature interactions, enabling the detection head to extract more delicate target features. Finally, an Inner-MPDIoU loss function is designed to accelerate model convergence and enhance target localization accuracy. Experimental results show that the proposed TINY-YOLOv8 algorithm outperforms other target detection methods. Compared with the benchmark model, the proposed approach improves mAP50 by 5.2% and 8% on the VisDrone and TinyPerson datasets, respectively, while reducing the number of parameters by 10.1%. These results indicate that the proposed algorithm is well-suited for aerial image target detection applications.