《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (5): 1658-1670.DOI: 10.11772/j.issn.1001-9081.2024050736

• 多媒体计算与计算机仿真 • 上一篇    

工业缺陷检测无监督深度学习方法综述

王文鹏, 秦寅畅, 师文轩()   

  1. 南开大学 软件学院,天津 300350
  • 收稿日期:2024-06-05 修回日期:2024-08-26 接受日期:2024-08-28 发布日期:2024-09-04 出版日期:2025-05-10
  • 通讯作者: 师文轩
  • 作者简介:王文鹏(2001—),男,山东德州人,硕士研究生,CCF会员,主要研究方向:大数据分析与应用
    秦寅畅(2000—),男,湖南长沙人,硕士研究生,CCF会员,主要研究方向:大数据分析与应用
    师文轩(1977—),男,河北廊坊人,副教授,博士,CCF会员,主要研究方向:机器学习、区块链。

Review of unsupervised deep learning methods for industrial defect detection

Wenpeng WANG, Yinchang QIN, Wenxuan SHI()   

  1. College of Software,Nankai University,Tianjin 300350,China
  • Received:2024-06-05 Revised:2024-08-26 Accepted:2024-08-28 Online:2024-09-04 Published:2025-05-10
  • Contact: Wenxuan SHI
  • About author:WANG Wenpeng, born in 2001, M.S. candidate. His research interests include big data analysis and applications.
    QIN Yinchang, born in 2000, M.S. candidate. His research interests include big data analysis and applications.
    SHI Wenxuan, born in 1977, Ph.D., associate professor. His research interests include machine learning, blockchain.

摘要:

工业缺陷检测在保障产品质量、提高企业竞争力方面具有极其重要的作用。传统的缺陷检测方法依赖人工检查,成本高且效率低下,难以满足大规模的质量检验需求。近年来,基于视觉的工业缺陷检测技术取得了显著进步,已成为产品外观质量检验的一种高效解决方案。但在许多实际工业场景中,获取大量带有标签的数据非常困难,且对产品检测的人工成本和实时性均有所要求,因此,无监督学习逐渐成为研究的热点。针对该领域任务构建、现行技术、评估标准以及不同方法之间的共性和差异,对相关工作进行综述。首先,明确工业缺陷问题的定义,并从数据难点和任务挑战等多个角度分析该问题的难点;其次,重点介绍基于无监督深度学习的工业缺陷检测主流方法,并对它们进行详细的归纳与分析;再次,介绍常用的公开数据集与评价指标;最后,对工业缺陷检测领域将来的工作进行展望。

关键词: 缺陷检测, 异常检测, 工业视觉, 深度学习, 无监督学习

Abstract:

Industrial defect detection plays a crucial role in ensuring product quality and enhancing enterprise competitiveness. Traditional defect detection methods rely on manual inspection, which is costly and inefficient, making it difficult to meet large-scale quality inspection requirements. In recent years, vision-based industrial defect detection technologies have made significant progress and become an efficient solution for product appearance quality inspection. However, in many practical industrial scenarios, it is challenging to obtain large amounts of labeled data, and there are requirements for both the labor cost and real-time performance of product detection, making unsupervised learning become a research hotspot. Related work on task construction, current technologies, evaluation standards, and the commonalities and differences among various methods in this field were reviewed. Firstly, the definition of industrial defect problems was clarified, and the difficulties of the problem were analyzed from perspectives of data challenges and task difficulties. Secondly, unsupervised deep learning-based methods for industrial defect detection were comprehensively introduced and systematically categorized. Furthermore, commonly used public datasets and evaluation metrics were summarized. Finally, future work in industrial defect detection was discussed.

Key words: defect detection, anomaly detection, industrial vision, deep learning, unsupervised learning

中图分类号: