A hierarchical training-convolutional neural network with feature alignment for steel surface defect recognition. (June 2023)

• Record Type: Journal Article
• Authors: Gao, Yiping; Gao, Liang; Li, Xinyu
• Is Part Of: Robotics and computer-integrated manufacturing. Volume 81(2023)
• Publication Date: 2023-06
• Subjects: Steel surface defect -- Convolutional neural network -- Feature alignment -- Hierarchical training; Robots, Industrial -- Periodicals; Computer integrated manufacturing systems -- Periodicals; Robotics -- Periodicals; Robots industriels -- Périodiques; Productique -- Périodiques; Robotique -- Périodiques; 670.285
• DOI: 10.1016/j.rcim.2022.102507
• Languages: English
• ISSNs: 0736-5845

Abstract

Highlights: Steel is an important material in modern manufacturing, and surface defect is one of the vital problems in steel production. A hierarchical training-convolutional neural network with feature alignment is proposed to improve the recognition results of steel surface defects by mapping the unrecognizable defect into the recognizable area. The proposed method has achieved a good performance on a public dataset with an accuracy of 100%. The proposed method has been developed into a real-world case successfully, and the accuracy is improved by 12.05%. Abstract: Steel is a basic material, and vision-based defect recognition is important for quality. Recently, deep learning, especially convolutional neural network (CNN), has become a research hotspot. However, steel defects have poor class separation, which is similar to the background, and different defects show similar textures. This causes some defects unrecognizable and influences production greatly. Thus, current CNNs still need to be improved. With this goal, this paper proposes a hierarchical training-CNN with feature alignment. The proposed method introduces a feature alignment, which maps the unrecognizable defects to the recognizable area, and a hierarchical training strategy is used to integrate the feature alignment into the training process. With these improvements, the proposed method achieves improved performance. The recognition results on a public dataset achieve 100%, which outperforms the other CNNs. And it has been developed into a real-world case successfully, which is significantly improved.