DOI：10.13705/j.issn.1671-6841.2023239

Review of Surface Defect Detection Methods Based on Few-shot Learning

CHEN Li ^1，2，3 ， YIN Xiangting ^1，2 ， JIN Qifan ^1，2 ， JIANG Xiaoheng1.²， JIU Mingyuan ^1，2 ， XU Mingliang ^1，2 （1.School of Computer and Artificial Intelligence， Zhengzhou University， Zhengzhou 45ooo1，China ;2. National Supercomputing Zhengzhou Center， Zhengzhou 450001， China; 3. School of PhysicalEducation （Main Campus）， Zhengzhou University， Zhengzhou 45ooO1， China）

Abstract： In some industrial scenarios，insufficient defect samples and labeling time-consuming and labor-intensive defects，limit the application of machine vision methods in surface defect detection.Technologies of industrial defect detection based on few-shot learning were introduced from three aspects ： image acquisition，image processing，and defect detection.Firstly，defect detection methods were divided into traditional surface defect detection methods and few-shot deep learning based defect detection methods.The traditional surface defect detection method was based on the manually extracted features to identify defects，which could be divided into three parts： defect segmentation，artificial feature extraction and defect recognition. Few-shot deep learning based industrial defect detection methods include data enhancement，transfer learning，model fine-tuning，semi-supervised learning，weakly supervised learning，unsupervised learning methods，etc. Secondly，some commonly used defect detection datasets and evaluation criteria of detection results were introduced.Finally，the existing problems and future research directions of few-shot learning based surface defect detection were discussed.

Key words： defect detection; few-shot learning；machine vision; deep learning

0 引言

在工业生产中，缺陷检测对于保证产品质量起着重要作用。工业品表面缺陷指工业制品与正常样本相比存在的缺失、瑕疵等，属于制品质量问题。常见的表面缺陷包括斑点、裂纹、凸起、划痕、砂眼等。这些缺陷不但影响制品外观，而且严重影响制品性能，降低制品使用寿命，甚至危及生命安全。表面缺陷检测是指对工业制品表面的划痕、异物屏蔽、颜色污染、孔洞等缺陷的检测，以获得制品表面缺陷的类别、轮廓、位置、尺寸等信息。目前常用的缺陷检测方法主要为人工检测，该方法虽具有灵活度高、技术门槛低、短期成本低等优势，但存在如下弊端：效率低，易受检测者主观因素和工作经验影响；当缺陷较小时，人眼难以识别，易发生漏检；有些检测环境对人体有害，不适于人工检测。基于视觉感知的工业品表面缺陷检测综合利用了数据处理、计算机视觉、机器学习等技术自动检测工件表面缺陷，具有效率高、性能稳定、不易受人眼物理条件和人类主观因素的影响等优势，可在复杂的生产环境中长时间工作。近年来，基于视觉感知的智能化缺陷检测已逐渐代替人工检测，成为工业品表面缺陷检测中的重要方法，广泛应用于瓷砖[1]、织物[2]、钢板[3-4]、印制电路板[5-6]等领域。

基于视觉感知的工业品缺陷检测可分为传统的基于图像处理的工业品缺陷检测和基于深度学习的工业品缺陷检测两种。前者需人工设计特征，避免了使用大量数据来学习，但对环境适应性较差，检测精度难以满足实际生产要求。……