To address issues such as missed and false detections of small targets, as well as the complexity of backgrounds in detecting dress code violations among power operators, we propose a novel algorithm, GMMI-RTDETR, for detecting dress code violations. First, we design a new backbone network to extract more detailed information, utilizing multiple attention mechanisms and gating mechanisms to fully capture the features of small targets. Next, we propose a feature aggregation module to enhance the model’s ability to perceive dress shapes in complex scenarios, while generating richer semantic information. Building upon this, we design a foreground enhancement network to improve the model’s detection capability of dress code violations in complicated settings. Finally, we integrate Inner-IoU and Powerful-IoU to form Inner-PIoU, which improves detection accuracy for small target and accelerates model convergence. Extensive experimental results on our private DEDV dataset and the publicly available SHWD dataset demonstrate that the proposed GMMI-RTDETR outperforms other models.