Abstract: For the problem of defects on steel surfaces, a defect detection algorithm is proposed to improve the YOLOv8n model. First, the backbone network of YOLOv8n is improved using a focused linear attention mechanism to enhance the model's ability to detect smaller targets in steel defects. Second, the model inference delay is reduced and the model performance is improved by introducing the SeaFormer module. Again, the CIoU loss function of the YOLOv8n model is converted to MPDIoU, which simplifies the computation process while taking into account the overlapping or non-overlapping regions, centroid distances, as well as width and height deviations. Finally, the detection accuracy is further improved by the knowledge distillation technique without increasing the model size and the number of parameters. The improved algorithm achieves a mAP50 of 77% on the GC10-DET dataset, which is 12.9% higher than the original YOLOv8n model. And the AP values reached 96.0%, 99.5%, 97.2% and 99.5% in the identification of punched holes, welded seams, creases and waist creases, respectively. The experimental results prove that the improved model has obvious advantages in the detection performance of steel surface defects detection.
高静, 刘宇阳, 郑力军. 基于改进YOLOv8n的钢材表面缺陷检测算法研究[J]. 吉林化工学院学报, 2026, 43(1): 8-15.
GAO Jing, LIU Yuyang, ZHENG Lijun. Research on Steel Surface Defect Detection Algorithm based on the Improved YOLOv8n Algorithm . Journal of Jilin Institute of Chemical Technology, 2026, 43(1): 8-15.
