Research on Surface Defect Detection and Intelligent Identification Method of Hot-Rolled Strip Based on Deep Learning

S.C. Xie, Y.X. An, H. Wang, J.T. Yang, Y.H. Lin, and G.Z. Ren

DOI: https://doi.org/10.64486/m.65.4.15
Online publication date: May 15, 2026

Abstract: Hot-rolled steel strips play a crucial role in industrial settings, where the accurate identification of surface defects is essential to uphold product quality and safety. This study introduces an enhanced version of the YOLOv8 model by integrating an Efficient Multi-scale Attention (EMA) mechanism into the C2f module, thereby creating the C2f_EMA module. This integration aims to improve the adaptive feature representation in both channel and spatial dimensions. The EMA mechanism serves to emphasize critical defect areas, suppress irrelevant background details, and enhance the detection precision of intricate and small defects. Evaluation on the NEU-DET dataset reveals that the upgraded model exhibits superior detection accuracy across most of the six defect categories, resulting in an overall mean average precision boost from 76.1 % to 77.6 %. Particularly noteworthy is the substantial enhancement in detecting small and medium-scale defects like Inclusion, Scratches, and Crazing. These findings underscore the efficacy of the C2f_EMA module in augmenting multi-scale feature representation within the YOLOv8 framework, all while preserving its lightweight nature and real-time performance. Consequently, this approach proves to be well-suited for surface defect identification in hot-rolled steel strip production lines.

Keywords: steel strip; hot-rolled; surface defect; object detection; YOLOv8; attention mechanism

This article is published online first and will appear in Metalurgija, Vol. 65, Issue 4 (2026).