Abstract
Surface defects are inevitable in the production of steel. However, traditional methods in industrial production face great challenges in detecting complex defects. Therefore, we propose LCED-YOLO based on YOLOv11 for steel defect detection. Firstly, an edge information enhancement module, C3K2-MSE, is designed to strengthen the extraction of edge information. Secondly, LDConv is introduced to lightweight the neck structure and reduce parameters. Then, a lightweight decoupling head designed for model detection tasks is proposed, further achieving model lightweighting. Finally, by introducing a learnable attention factor to optimize the CIoU loss, we focused on locating difficult samples, enhancing the detection capability. A large number of experiments were conducted on the NEU-DET and GC10-DET datasets. Compared to YOLOv11, the mAP50 of the proposed model improved by 2.6% and 3.3%, attaining 79.8% and 70.3%, respectively. It decreased 19% of parameters and 23% of floating-point operations, fulfilling the needs of lightweight and detection precision.