NA-YOLO: an enhanced anti-UAV detection network for complex scenarios based on YOLO11

The issues about security and privacy caused by the misuse of unmanned aerial vehicles (UAVs) have attracted increasing attention. Rapid and accurate UAV detection in complex scenarios has become critical for anti-UAV missions. This paper proposes an enhanced model for UAV detection in complex conditions based on YOLO11, aiming to achieve accurate and efficient detection under challenging conditions. Firstly, a small-object detection head is introduced. The large-object detection head is removed. This design improves sensitivity to tiny targets. It also reduces model parameters and increases inference speed. Then, a feature fusion strategy based on FPN is employed to prevent the loss of low-level features that may occur during multi-scale fusion. In addition, the WIoU v3 loss function is incorporated to mitigate oscillations that arise during the regression of small targets. Finally, a feature enhancement module termed C3k2-NAM is proposed. It integrates a normalization-based attention mechanism into the C3k2 block. This design improves the discriminative capability of features for UAV targets. At the same time, it suppresses background interference in complex conditions. The proposed method achieves a comparable inference speed to the baseline model on the DUT Anti-UAV dataset. Precision, recall, mAP0.5, and mAP@0.5 :0.95 increase by 10%, 11%, 8.3%, and 24.3%, respectively. These results demonstrate significant overall performance gains. The improvement in mAP@0.5 :0.95 is particularly notable. Furthermore, the proposed method enables more accurate UAV detection under stricter IoU thresholds in complex environments.