Scholarlink

<p>Textile tow defect detection faces challenges such as slow speed, small object sizes, and low accuracy. To address these issues, we propose PRD-YOLOv11, a model that improves accuracy while maintaining speed. A key innovation is the Progressive Feature Compression Downsampling (PFCD) method, which uses layered compression and decompression to reduce parameters and memory usage while preserving semantic information. PFCD includes two versions: Complex PFCD (CPFCD) for shallow feature extraction and Simple PFCD (SPFCD) for deep feature extraction, balancing efficiency and representation. Additionally, the Region-Guided Dynamic Attention (RGA) mechanism enhances feature extraction by adaptively allocating weights and aggregating context features, improving robustness in complex backgrounds. We also introduce Dual-Stream Loss (DS) to accelerate convergence and enhance detection capabilities beyond traditional IoU metrics. Experiments on our custom Cellulose Tow (CT) dataset show PRD-YOLOv11 achieves an mAP of 69.3% (a 4.7% improvement) and F1 score of 70.2% (a 6.6% improvement), with only a 0.3 GFLOPs increase, and a real-time detection speed of 47 FPS. To further validate the model’s generalization ability, we tested it on public datasets NEU, it achieves mAP scores of 80.7%, demonstrating superior performance.</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

PRD-YOLOv11: Efficient and Accurate Textile Tow Defect Detection via Progressive Representation Distillation

  • 期刊论文

  • 发表日期：

    2025年5月29日

查看全文
  • Peng Chen,
  • YuGang Luo,
  • Jun Zhang,
  • Bing Wang

摘要

Textile tow defect detection faces challenges such as slow speed, small object sizes, and low accuracy. To address these issues, we propose PRD-YOLOv11, a model that improves accuracy while maintaining speed. A key innovation is the Progressive Feature Compression Downsampling (PFCD) method, which uses layered compression and decompression to reduce parameters and memory usage while preserving semantic information. PFCD includes two versions: Complex PFCD (CPFCD) for shallow feature extraction and Simple PFCD (SPFCD) for deep feature extraction, balancing efficiency and representation. Additionally, the Region-Guided Dynamic Attention (RGA) mechanism enhances feature extraction by adaptively allocating weights and aggregating context features, improving robustness in complex backgrounds. We also introduce Dual-Stream Loss (DS) to accelerate convergence and enhance detection capabilities beyond traditional IoU metrics. Experiments on our custom Cellulose Tow (CT) dataset show PRD-YOLOv11 achieves an mAP of 69.3% (a 4.7% improvement) and F1 score of 70.2% (a 6.6% improvement), with only a 0.3 GFLOPs increase, and a real-time detection speed of 47 FPS. To further validate the model’s generalization ability, we tested it on public datasets NEU, it achieves mAP scores of 80.7%, demonstrating superior performance.