EDS-YOLO: an enhanced steel surface defect detection method based on improved YOLO11

Abstract Steel surface defect detection plays a critical role in quality inspection and process control during steel manufacturing. However, in real industrial environments, defects often exhibit low contrast, blurred boundaries, and significant variations in scale and morphology, which continuously challenge the reliability and robustness of detection models. To address these issues, this paper proposes an enhanced steel surface defect detection model, termed EDS-YOLO, based on YOLO11n. Specifically, an edge-spatial information fusion module is designed to explicitly integrate edge cues with spatial features, thereby effectively enhancing the representation of defects with blurred boundaries and subtle characteristics. In addition, a dilated shared convolution feature pyramid is introduced, which employs a parameter-sharing strategy across multi-scale dilated convolutions to efficiently model defect patterns at different scales with low computational overhead. Furthermore, a Slim-neck structure integrating lightweight Ghost Shuffle Convolution and VoV-GSCSP modules is incorporated to reduce feature redundancy while maintaining efficient and effective feature fusion. Experimental results on the public NEU-DET and GC10-DET datasets demonstrate that the proposed EDS-YOLO improves the mAP50 by 4.2 % and 4.4 % , respectively, compared with the baseline model. Moreover, EDS-YOLO consistently exhibits stable advantages in detection accuracy, computational efficiency, and generalization capability, highlighting its strong potential for practical industrial applications.