An improved method for surface defect detection in hot-rolled strip-steel

カテゴリ: 研究会(論文単位)

論文No: CMN23022

グループ名: 【C】電子・情報・システム部門 通信研究会

発行日: 2023/03/26

タイトル(英語): An improved method for surface defect detection in hot-rolled strip-steel

著者名: BISWAS SOUMEN(HITACHI INDIA PVT. LTD.),GANESH ANANTH(HITACHI INDIA PVT. LTD.)

著者名(英語): SOUMEN BISWAS(HITACHI INDIA PVT. LTD.),ANANTH GANESH(HITACHI INDIA PVT. LTD.)

キーワード: Defect Detection|Steel Surface|YOLOv5|Improved YOLOv5 |Defect Detection|Steel Surface|YOLOv5|Improved YOLOv5

要約(日本語): Computer vision (CV) techniques play a pivotal role in industrial appliances. Developing various intelligent processes to solve real-world problems has gradually become essential in today’s world. The critical aspect is that implementing artificial intelligence (AI) based CV models in various industrial sections, especially in production units, is a challenging task because defects in the product increase the production cost and reduce the product quality. Thus, the application of object detection to find defects is significant today in industrial units. Steel is one of the raw materials in the industry, with various defects in steel surfaces. The steel is in the hot-rolled strip, where defect detection is vital to increase production quality. The hot-rolled strip has multiple defects, which are difficult to identify with bare eyes. Hence, manual inspection often led to either false or slow detection of defects in the hot-rolled strip. Defect detection accuracy plays a vital role in detecting defects and improving production speed. Thus, an automated defect detection process in steel surfaces yields better results than a manual inspection process. CV techniques come up with various defect detection strategies, which resolve most of the challenges yet improving to solve many challenges in industrial aspects. _x000D_ _x000D_ The present research proposes a steel surface defect detection method where we have improved the baseline YOLOv5 architecture. In this work, the improved YOLOv5 model has three different parts, i.e., backbone, head, and output. In the proposed model, we have changed the network both in backbone and head. In backbone of our improved YOLOv5, we have used depth wise separable convolution layer and cross convolution layer with C3 module. The head of the improved YOLOv5 is a novel architecture utilizing the ghost convolution layer and ghost bottleneck network. The overall architecture gives three anchors detection output results. The above changes in the improved YOLOv5 reduce the model size, computation complexity and floating-point operations (FLOPS). The improved YOLOv5 also increases the performance compared to the baseline YOLOv5 model. We considered the NEU-DET dataset, which consists of six different types of defects in hot-rolled strip. The experiment shows an improvement in real-time defect detection accuracy with test set images. The average detection accuracy using the improved YOLOv5 model increases by 4% compared to the baseline YOLOv5 model. Further experiments in NEU-DET dataset using YOLOv3, faster RCNN, and RetinaNet provides mAP (mean average precision) of 68%, 71%, and 72% for hot-rolled defect detection. The comparison of mAP using improved YOLOv5 shows an improvement of 8%, 5%, and 4% for defect detection over YOLOv3, faster RCNN, and RetinaNet. Also, the improved YOLOv5 reduces the FLOPS compared to existing baseline models. These improvements significantly improve the defect detection in hot-rolled strip steel._x000D_

要約(英語): Computer vision (CV) techniques play a pivotal role in industrial appliances. Developing various intelligent processes to solve real-world problems has gradually become essential in today’s world. The critical aspect is that implementing artificial intelligence (AI) based CV models in various industrial sections, especially in production units, is a challenging task because defects in the product increase the production cost and reduce the product quality. Thus, the application of object detection to find defects is significant today in industrial units. Steel is one of the raw materials in the industry, with various defects in steel surfaces. The steel is in the hot-rolled strip, where defect detection is vital to increase production quality. The hot-rolled strip has multiple defects, which are difficult to identify with bare eyes. Hence, manual inspection often led to either false or slow detection of defects in the hot-rolled strip. Defect detection accuracy plays a vital role in detecting defects and improving production speed. Thus, an automated defect detection process in steel surfaces yields better results than a manual inspection process. CV techniques come up with various defect detection strategies, which resolve most of the challenges yet improving to solve many challenges in industrial aspects. _x000D_ _x000D_ The present research proposes a steel surface defect detection method where we have improved the baseline YOLOv5 architecture. In this work, the improved YOLOv5 model has three different parts, i.e., backbone, head, and output. In the proposed model, we have changed the network both in backbone and head. In backbone of our improved YOLOv5, we have used depth wise separable convolution layer and cross convolution layer with C3 module. The head of the improved YOLOv5 is a novel architecture utilizing the ghost convolution layer and ghost bottleneck network. The overall architecture gives three anchors detection output results. The above changes in the improved YOLOv5 reduce the model size, computation complexity and floating-point operations (FLOPS). The improved YOLOv5 also increases the performance compared to the baseline YOLOv5 model. We considered the NEU-DET dataset, which consists of six different types of defects in hot-rolled strip. The experiment shows an improvement in real-time defect detection accuracy with test set images. The average detection accuracy using the improved YOLOv5 model increases by 4% compared to the baseline YOLOv5 model. Further experiments in NEU-DET dataset using YOLOv3, faster RCNN, and RetinaNet provides mAP (mean average precision) of 68%, 71%, and 72% for hot-rolled defect detection. The comparison of mAP using improved YOLOv5 shows an improvement of 8%, 5%, and 4% for defect detection over YOLOv3, faster RCNN, and RetinaNet. Also, the improved YOLOv5 reduces the FLOPS compared to existing baseline models. These improvements significantly improve the defect detection in hot-rolled strip steel._x000D_

本誌: 2023年3月29日通信研究会

本誌掲載ページ: 7-11 p

原稿種別: 英語

PDFファイルサイズ: 1,510 Kバイト