MAFD-NET: multi-head attention dynamic sampling fusion network for power tower detection in high-resolution remote sensing images

With the rapid development of high-resolution remote sensing technology, power tower detection from remote sensing imagery has emerged as a critical research direction for intelligent power grid inspection. However, existing deep learning models still struggle with three key challenges in high-resolution scenes: poor adaptability to diverse power tower scales, severe feature confusion induced by complex backgrounds, and high miss rates for small-scale distribution towers. To alleviate these issues, this paper proposes a multi-head attention dynamic sampling fusion network (MAFD-NET) based on YOLO11. A multi-head attention feature fusion (MAFB) module is embedded in the backbone to strengthen global and local feature extraction by integrating the Mamba mechanism and hierarchical multi-head self-attention. Meanwhile, an Edge-Preserving Dynamic Upsampling (EPDU) module is designed in the neck to achieve accurate alignment of multi-scale features. Experiments are carried out on the Duke Power Transmission and Distribution Infrastructure dataset and the Aerial-Power dataset with standard preprocessing. Results show that the proposed method achieves 78.26% mAP 0.5 , 75.12% recall, and 74.87% mIoU on the Duke dataset, and 76.18% mAP 0.5 , 74.32% recall, and 71.83% mIoU on the Aerial-Power dataset. All metrics exceed those of mainstream detectors, significantly improving the accuracy and robustness of power tower detection in complex scenes. The method provides reliable technical support for intelligent inspection of power facilities.