现代纺织技术 ›› 2023, Vol. 31 ›› Issue (4): 155-163.

• • 上一篇    下一篇

基于改进YOLOv5模型的织物疵点检测

  

  1. 1. 武汉纺织大学纺织科学与工程学院,武汉 430200;2.中电建湖北电力建设有限公司, 武汉430040
  • 收稿日期:2022-09-27 出版日期:2023-07-10 网络出版日期:2023-09-12
  • 通讯作者:
  • 作者简介:高敏(1996—),女,江西九江人,硕士研究生,主要从事图像处理方面的研究。
  • 基金资助:
    国家自然科学基金青年基金项目(51503162);湖北省自然科学基金青年面上项目(2016CFB459);湖北省大学生创新训练计划项目(S201910495063);国家大学生创新训练计划项目(201910495014)

Fabric defect detection based on improved YOLOv5 model

  1. 1. School of Textile Science and Engineering, Wuhan Textile University, Wuhan 430200, China; 2. PowerChina Hubei Engineering Co., Ltd., Wuhan 430040, China
  • Received:2022-09-27 Published:2023-07-10 Online:2023-09-12

摘要: 针对传统机器学习方法检测织物疵点精度低,小目标检测较困难的问题,提出一种基于改进YOLOv5的织物疵点的目标检测算法。在YOLOv5模型的Backbone模块中分别引入SE注意力机制和CBAM注意力机制,使模型聚焦于图像中的关键信息,改进传统 YOLOv5网络检测精度不高的问题。结果表明:改进后的模型具有更好的检测性能,其中引入CBAM模块后提升幅度最明显,较原网络mAP值提升了7.7%,基本满足织物疵点检测需求。

关键词: 织物疵点, YOLOv5模型, 注意力机制, 深度学习

Abstract: China is the largest producer and consumer of textiles in the world, and the textile industry is a traditional industry in China. The industry affects people's life and employment issues, and plays an important role in the national economy. The widely used textile products are indispensable in people's lives and the quality of them not only affects people's lives, but the development of textile enterprises. Therefore, in the production process of textile products, the quality inspection is a very important link in the production chain.
For a long time, due to the limitations of the cloth making process and production equipment, the surface of the cloth is often stained, and there are broken figures and other defects. Traditional detection methods have many disadvantages. On the one hand, the detection efficiency of cloth inspection workers is relatively low, and a cloth inspector can find 200 defects per hour at most. Because some defects are small and difficult to be found, the defect detection rate is only 70%, and the concentration lasts half an hour at most. If this time range is exceeded, visual fatigue will occur. As workers rely on subjective experience when inspecting the quality of cloth, and long-term work will cause visual fatigue, missed inspections and wrong inspections will inevitably occur, leading to low defect detection efficiency of cloth. On the other hand, because the cloth inspection work requires a lot of eyesight and energy of the workers, it is difficult to recruit workers and the cost is high. If there is a missed inspection in the cloth inspection process and the defects are not picked out, the products produced will be classified as defective products even if the cloth has passed the follow-up cumbersome processing, which will cause certain economic losses to the enterprise.
With the development of artificial intelligence, all walks of life are moving towards intelligent production. Computer vision has replaced humans eyes and brains to observe the world through cameras and computers, automatically analyzing the collected videos or pictures. In the field of target detection, with the rise of deep learning, the performance of target detection has also been greatly improved.
To address such problems as unsatisfactory detection effects for small targets, we conducted research on the surface defects of cloth by combining with the deep learning network, proposing an algorithm model based on deep learning for defect detection and classification. This model can improve the detection accuracy and reduce the missed detection rate of small targets, which meets the needs of actual production for detection of cloth defects, and has certain practical significance on the intelligent development of the weaving industry.

Key words: fabric defects, YOLOv5 model, attention mechanism, deep learning

中图分类号: