High-performance RT-DETR model for industrial defect detection with semantic guidance and hierarchical attention

by Jin Huang, Bin Gu, Yunlong Tang, Bingyue Xu, Puxuan Li, Jiayu Ye, Jingwen Xu, Taihua Zhang

With the advancement of intelligent manufacturing and Industry 4.0, surface defect detection plays a critical role in ensuring product quality and production safety. To address the limitations of existing detection models in handling small sample sizes, complex textures, and multi-scale defects, this paper proposes a high-performance industrial defect detection model based on the RT-DETR framework, incorporating semantic guidance and hierarchical attention mechanisms. Specifically, a Semantic-Guided Query Enhancement Module is designed to strengthen the contextual awareness of query vectors through multi-source semantic paths and a residual feedback structure. Additionally, a Hierarchical Attention Fusion Structure is constructed to build interactive graphs among multi-scale features, achieving cross-scale semantic alignment and structural consistency modeling. Experiments conducted on three benchmark industrial defect datasets—NEU-DET, DAGM2007, and PCB-DET—demonstrate the effectiveness of the proposed method, achieving mAP@0.5 scores of 78.9%, 84.7%, and 87.4%, respectively, outperforming the best baseline models by 1.2% to 3.0%. For the more stringent mAP@0.5:0.95 metric, the method achieves 44.3%, 48.1%, and 52.3%, significantly surpassing mainstream models such as YOLOv8, and BMA-YOLO. Furthermore, Grad-CAM visualizations validate the model’s superior focus capability and boundary-fitting accuracy in regions with complex textures and sparse targets. Overall, the proposed architecture enhances semantic perception, scale robustness, and generalization performance in industrial defect detection while maintaining real-time efficiency.