YOLO12 and YOLO13 for Steel Surface Defect Detection: Experimental Evaluation

Abstract

Intelligent inspection and defect detection of steel surfaces are essential for ensuring product quality, minimizing production losses, and supporting quality control in modern manufacturing systems. In particular, the identification of defects on steel plates is an important step, as surface imperfections can significantly affect the mechanical performance and commercial value of the final product. This study presents a comparative evaluation of YOLO12 and YOLO13 architectures for steel surface defect detection using the NEU-DET dataset under different validation ratios and input image resolutions. Four models (YOLO12-S, YOLO12-L, YOLO13-S, and YOLO13-L) were evaluated under three experimental settings using standard object detection metrics and three different random seeds. Experimental results showed that YOLO13-S consistently demonstrated superior overall detection performance, reaching a best single-run mAP@50 value of 0.753 with seed 42 and a precision value of 0.848 with seed 44. In addition, it achieved the highest single-run mAP@50–95 value of 0.449, with seed 43. In contrast, YOLO12-L achieved the highest single-run recall (0.749), with seed 42, indicating its stronger capability for detecting positive defect instances. Overall, the results demonstrate that YOLO13-S provides the best balance between detection accuracy and computational efficiency, while YOLO12-L offers superior recall performance under the evaluated conditions.