Scholarlink

<p>In fabric production, an efficient inspection system for fabric rolls before garment manufacturing is crucial. The system must cover large areas of moving fabric to reduce hardware costs while accurately detecting defects. Challenges include identifying small, blurry, complex, or overlapping defects during fast fabric movement, which can lead to missed issues and compromised quality. In this paper, we present a fabric defect detection system combining an optimized optical design with a deep learning model. The system uses dark field (DF) and back light (BL) techniques for rapid and precise detection of subtle defects during the re-rolling process. Data is collected through the designed optical system, and diverse data sources are generated using various image augmentation techniques, resulting in a dataset of 3355 images. The dataset is trained on three models: RefineDet, improved RefineDet, Faster R-CNN, DETR, YOLOv8, and YOLOv11. The experimental results demonstrate that the Faster R-CNN, YOLOv8, and YOLOv11 models outperform other models in terms of accuracy, achieving mAP scores of 86% and 81%, respectively. Additionally, YOLOv8 and YOLOv11 are more compact, enabling them to process recognition tasks twice as fast as Faster R-CNN. With its enhanced architecture, YOLOv11 stands out in its ability to accurately recognize small, distant, and blurry objects while maintaining the speed necessary for real-time recognition system deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated fabric defect classification in textile manufacturing using advanced optical and deep learning techniques

  • 期刊论文

  • 发表日期：

    2025年3月12日

查看全文
  • Hieu Dau Sy,
  • Phuc Dang Thi,
  • Huy Vo Gia,
  • Khang Le Nguyen An

摘要

In fabric production, an efficient inspection system for fabric rolls before garment manufacturing is crucial. The system must cover large areas of moving fabric to reduce hardware costs while accurately detecting defects. Challenges include identifying small, blurry, complex, or overlapping defects during fast fabric movement, which can lead to missed issues and compromised quality. In this paper, we present a fabric defect detection system combining an optimized optical design with a deep learning model. The system uses dark field (DF) and back light (BL) techniques for rapid and precise detection of subtle defects during the re-rolling process. Data is collected through the designed optical system, and diverse data sources are generated using various image augmentation techniques, resulting in a dataset of 3355 images. The dataset is trained on three models: RefineDet, improved RefineDet, Faster R-CNN, DETR, YOLOv8, and YOLOv11. The experimental results demonstrate that the Faster R-CNN, YOLOv8, and YOLOv11 models outperform other models in terms of accuracy, achieving mAP scores of 86% and 81%, respectively. Additionally, YOLOv8 and YOLOv11 are more compact, enabling them to process recognition tasks twice as fast as Faster R-CNN. With its enhanced architecture, YOLOv11 stands out in its ability to accurately recognize small, distant, and blurry objects while maintaining the speed necessary for real-time recognition system deployment.