SCDF-YOLOv11: Selection and compensation detailed feature integration network for UAV detection

To address the limitations of the existing object detection algorithms in detecting Unmanned Aerial Vehicles (UAVs), such as weak detection capabilities and high computational complexity, a Selection and Compensation Detailed Feature integration network based on YOLOv11 (SCDF-YOLO) that balances lightweight design and accuracy is proposed. First, a Parallel Branch Feature Extraction (PBFE) structure is designed. It utilizes feature extraction auxiliary branches to focus on detailed information, which is beneficial to enriching the features of the backbone branch. Second, a Selection and Compensation Feature Integration (SCFI) mechanism is proposed, and the semantic information from high-level feature maps guides the selection of high-quality features from low-level feature maps. To highlight object details and expand the difference between similar objects, saliency information is then introduced to compensate for the fused features. Finally, the SCFI-Neck is constructed based on SCFI, which integrates multi-scale features to improve the detection accuracy of UAVs in complex backgrounds. Experimental results demonstrate that compared to YOLOv11n, the proposed SCDF-YOLO maintains a lightweight model while significantly improving detection accuracy on both the Anti-UAV-dataset testing set and the UAV-bamboo-dataset testing set. The further optimized SCDF-YOLO-S shows even better performance, with a 6.6% increase in mAP50 and a 5.7% increase in mAP50:95 on the Anti-UAV-dataset testing set. Similarly, on the UAV-bamboo-dataset testing set, mAP50 improves by 3.9% and mAP50:95 improves by 7%. These improvements enable the model to meet the high-accuracy UAV detection requirements in complex scenarios.