RCCF-YOLO: region-constrained channel weighting and multi-scale fusion network for power tower detection in high-resolution SAR images

To address the issue of frequent false alarms in transmission tower detection in SAR images, this paper proposes a multi-task detection framework named RCCF-YOLO, which incorporates regional constraints. The model employs a unified feature extraction network and constructs dual branches for object recognition and scene classification, enabling joint learning of object-level and region-level features. Based on the lightweight DSC3k2 module, the model introduces the DAE module (channel diagonal attention enhancement block), which models channel importance through channel diagonal weighting and dynamic attention, enhancing the discriminative capability for small-scale targets. The results show that RCCF-YOLO achieves a precision of 0.940 and a recall of 0.880 on the test set, with an mAP50 of 0.934, representing a 2.9% improvement over YOLOv13. In field tests conducted in the Three Gorges area, the proposed model significantly reduces missed and false detections in dense tower scenarios, validating the effectiveness and robustness of the proposed method.