Abstract: Texture surface defect detection method plays a key role in the inspection of industrial products.It can detect the texture defects on the surface of various objects quickly and accurately,thereby improving product quality and production efficiency.In recent years,with the rapid development of deep learning technology,more and more texture surface defect detection methods based on deep learning have emerged,which have shown significant advantages in handling complex texture problems and achieving high-precision detection.Therefore,the texture surface defect detection methods based on deep learning in the past five years were systematically and comprehensively summarized from the following three aspects:the supervised learning methods,the weakly supervised learning methods and unsupervised learning methods.The latest developments in each category were concluded,and the methods were analyzed through performance comparisons on specific datasets to reveal current challenges and limitations.Finally,major challenges in this field were highlighted,and potential solutions were proposed.

Key words: texture surface defect detection, defect detection, deep learning, supervised learning, weakly supervised learning, unsupervised learning

摘要： 纹理表面缺陷检测方法在工业产品检测中发挥着关键作用,可以快速准确地检测出各种纹理表面上存在的缺陷,从而提升产品质量和生产效率。近年来,随着深度学习技术的快速发展,涌现出了越来越多的基于深度学习的纹理表面缺陷检测方法,这些方法在处理复杂纹理问题和实现高精度检测方面展现出了显著优势。因此,从监督学习、弱监督学习和无监督学习3个角度,系统、全面地总结了近5年来基于深度学习的纹理表面缺陷检测方法,归纳了各类方法的前沿成果,并通过在特定数据集上的性能对比,深刻剖析了各类方法面临的挑战与存在的不足。最后,总结了该领域面临的主要挑战,并提出了可能的解决策略。

关键词: 纹理表面缺陷, 缺陷检测, 深度学习, 监督学习, 弱监督学习, 无监督学习