Scholarlink

<p>This study proposed a lightweight deep-learning approach to localize loop-formation defects in air-jet textured yarn (ATY) directly from grayscale surface-loop images. Inspection of ATY is challenging due to its elongated geometry, semi-transparent filaments, and defect morphologies that differ from conventionally spun yarns. To preserve longitudinal context, we designed a structure-preserving convolutional neural network (CNN) that processes full-length images without cropping. Ground-truth defects—defined as loop-formation failure and insufficient core-effect yarn entanglement—were manually annotated using computer vision annotation tool (CVAT). Pre-processing with binarization and aspect-ratio preservation reduces noise and computational cost, while loop-density weighting increases sensitivity to defect-prone regions. Instead of dense masks, the network performs one-dimensional boundary regression, outputting two horizontal coordinates (<i>x</i><sub><i>1</i></sub><i>, x</i><sub><i>2</i></sub>) that delimit the defective span along the yarn axis, improving stability, and reducing complexity. Trained on 53 annotated images with data augmentation, the model was evaluated using mean absolute error (MAE), intersection over union (IoU), and expert visual inspection; MAE stabilized at approximately 130 epochs, and the mean IoU reached 0.41. Despite the limited dataset, targeted data refinement and the boundary-regression formulation produced accurate, interpretable localization at low computational cost. The method is suitable for ATY quality control and is potentially extensible to other filament-based yarns exhibiting similar defect morphologies.</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

Lightweight Deep-Learning for Defect Localization in Air-Jet Textured Yarn from Grayscale Surface-Loop Images

  • 期刊论文

  • 发表日期：

    2025年11月26日

查看全文
  • Hyeongmin Moon,
  • Md Morshedur Rahman,
  • Eunho Shin,
  • Ingi Hong,
  • Seunga Choi,
  • Hyungsup Kim,
  • Chang Kyu Park,
  • Joonseok Koh

摘要

This study proposed a lightweight deep-learning approach to localize loop-formation defects in air-jet textured yarn (ATY) directly from grayscale surface-loop images. Inspection of ATY is challenging due to its elongated geometry, semi-transparent filaments, and defect morphologies that differ from conventionally spun yarns. To preserve longitudinal context, we designed a structure-preserving convolutional neural network (CNN) that processes full-length images without cropping. Ground-truth defects—defined as loop-formation failure and insufficient core-effect yarn entanglement—were manually annotated using computer vision annotation tool (CVAT). Pre-processing with binarization and aspect-ratio preservation reduces noise and computational cost, while loop-density weighting increases sensitivity to defect-prone regions. Instead of dense masks, the network performs one-dimensional boundary regression, outputting two horizontal coordinates (x1, x2) that delimit the defective span along the yarn axis, improving stability, and reducing complexity. Trained on 53 annotated images with data augmentation, the model was evaluated using mean absolute error (MAE), intersection over union (IoU), and expert visual inspection; MAE stabilized at approximately 130 epochs, and the mean IoU reached 0.41. Despite the limited dataset, targeted data refinement and the boundary-regression formulation produced accurate, interpretable localization at low computational cost. The method is suitable for ATY quality control and is potentially extensible to other filament-based yarns exhibiting similar defect morphologies.