An Adaptive Shunt Model for Steel Defect Detection based on YOLOX

Defect detection plays a critical role in steel production. Detection methods based on deep learning have made remarkable progress in defect inspection. However, most detectors cannot handle the differentiated semantic hierarchies of defects in images and the dynamic changes of the model during training. Aiming at the above problems, the model ST-YOLO for steel defect detection is proposed in this paper, which adopts the shunt fusion network structure and the self-tuning label assignment algorithm. The network structure is designed for feature fusion to meet the computational requirements of the classification and localization tasks specifically. The self-tuning optimal transport assignment is proposed for label assignment to guide the model to accomplish training flexibly. Our model achieves 80.3 mAP at 46.0 FPS on the NEU-DET dataset, and it is verified to show excellent performance in steel defect detection applications.