Scholarlink

<p>Defect detection in complex textiles is a key challenge. Contrastive language-image pre-training (CLIP) performs well but is restricted by its single-text prompt from simultaneous segmentation and classification. A textile defect detection method based on collaborative fusion of dual-text prompts for CLIP (CFDT-CLIP) is proposed to integrate pixel-level and image-level defect detection effectively. A pixel-prioritized module (PPM) focuses on pixel-level segmentation, with a text data augmentation (TDA) module enhancing the diversity of feature texts containing defect information. An image-priority module (IPM) assigns a learnable generic text prompt to each image (ignoring specific defect categories) to differentiate defective/good images and prioritize image-level classification. Via the mixed output module (MOM), a learnable collaborative fusion matrix (LCFM) integrates and normalizes anomaly scores/maps from PPM and IPM for final results. Tested on textile datasets, the algorithm outperforms mainstream CLIP models in both tasks. The dual-branch PPM/IPM provides technical underpinnings for shortening inference time, enabling high-performance parallel computing, and empowering supercomputing.</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

CFDT-CLIP:fabric anomaly detection with collaborative fusion under dual-text prompts

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Junjie Zhuang,
  • Hao Liu,
  • Jiuzhen Liang,
  • Hao Wu

摘要

Defect detection in complex textiles is a key challenge. Contrastive language-image pre-training (CLIP) performs well but is restricted by its single-text prompt from simultaneous segmentation and classification. A textile defect detection method based on collaborative fusion of dual-text prompts for CLIP (CFDT-CLIP) is proposed to integrate pixel-level and image-level defect detection effectively. A pixel-prioritized module (PPM) focuses on pixel-level segmentation, with a text data augmentation (TDA) module enhancing the diversity of feature texts containing defect information. An image-priority module (IPM) assigns a learnable generic text prompt to each image (ignoring specific defect categories) to differentiate defective/good images and prioritize image-level classification. Via the mixed output module (MOM), a learnable collaborative fusion matrix (LCFM) integrates and normalizes anomaly scores/maps from PPM and IPM for final results. Tested on textile datasets, the algorithm outperforms mainstream CLIP models in both tasks. The dual-branch PPM/IPM provides technical underpinnings for shortening inference time, enabling high-performance parallel computing, and empowering supercomputing.