FGL-YOLOv8: A Fine-Grained Lightweight Algorithm for Small Object Detection in UAV Images

To enhance the precision of small object detection in UAV images and minimize missed detections, this paper proposes a fine-grained lightweight YOLOv8 algorithm for complex backgrounds and densely distributed small targets. Specifically, firstly, We introduce a small object detection head from high-resolution feature maps, effectively reducing the missed detection rate of small targets. Secondly, the model’s parameters are decreased by cropping the feature maps of the large object detection layer. Furthermore, we improve the model’s feature extraction capability for small objects by incorporating the Space-to-depth (SPD) concept into the backbone network. Experimental verification on the VisDrone dataset demonstrates significant performance improvements over the original model. Our approach achieves an 8.8% mAP50 improvement while reducing parameters by 62.39%. It achieves an FPS exceeding 74. Therefore, our method better satisfies the requirements for object detection in real-world drone image scenarios.