To address the challenges in detecting violations in power operators’ dress, such as missed and false detections of small objects, complex backgrounds, and inconsistent target scales, we propose a novel detection method, SDMD-RTDETR. First, we design the Small-Target Awareness Feature Extraction Module (SAFEM), which enhances small-target detection by extracting local features, global features, and surrounding contextual information. Second, we integrate Deformable Attention into the AIFI module, forming D-AIFI. This modification enables the model to focus attention more effectively on relevant dress features while reducing interference from irrelevant backgrounds, thereby improving detection performance in complex scenarios. Additionally, we design a Dynamic Multi-Scale Feature Fusion Network (DMFFN) to capture richer multi-scale information and apply weighted fusion to different-scale features, further enhancing the model’s detection accuracy for dress at different scales. Experimental results on our proprietary DEDV dataset show that SDMD-RTDETR improves mAP@0.5 and mAP@0.5:0.95 by 3.5 and 3.7 points, respectively, compared to the baseline RT-DETR. Extensive experiments on the publicly available SHWD dataset further validate the effectiveness of SDMD-RTDETR.