Surface defect detection on rolling mill rolls: a transfer learning approach with improved YOLOv5

Abstract The roller is a crucial tool in the hot-rolled strip steel manufacturing process. Rapid and accurate detection of roll surface defects is essential for enhancing product surface quality, ensuring dimensional precision, and reducing scrap rates. Currently, industrial roll inspection primaril...

Full description

Saved in:
Bibliographic Details
Main Authors: Yulong Hu, Wen Peng, Shiyi Chen, Jinyun Liu, Xudong Li, Jie Sun, Dianhua Zhang
Format: Article
Language:English
Published: Springer 2025-06-01
Series:Discover Applied Sciences
Subjects:
Online Access:https://doi.org/10.1007/s42452-025-06920-7
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract The roller is a crucial tool in the hot-rolled strip steel manufacturing process. Rapid and accurate detection of roll surface defects is essential for enhancing product surface quality, ensuring dimensional precision, and reducing scrap rates. Currently, industrial roll inspection primarily relies on manual visual assessment, which is prone to subjectivity, low accuracy, and inefficiency. To overcome these limitations, this study introduces a defect detection method based on transfer learning and an enhanced YOLOv5 model.Given the lack of publicly available datasets for roll surface defects, a representative dataset was first constructed using defect images collected from a steel production line. To further optimize the detection model, the NEU-DET dataset—containing strip surface defect images—was employed to pre-train the improved YOLOv5 model, refining its parameters. The pre-trained model was then adapted for roll defect detection using transfer learning, where the learned parameters from strip surface defects served as initial weights for training on the roll defect dataset. Experimental results demonstrate that the proposed TR-CNF-YOLOv5 model, integrating transfer learning with an improved YOLOv5 architecture, outperforms existing models. Specifically, it achieves an mAP improvement of approximately 9.8% over the original YOLOv5 and 7.1% over YOLOv7 on the roll surface defect dataset.
ISSN:3004-9261