欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (3): 377-386.DOI: 10.11996/JG.j.2095-302X.2022030377

• 图像处理与计算机视觉 • 上一篇    下一篇

基于异常检测的产品表面缺陷检测与分割

  

  1. 1. 华北电力大学控制与计算机工程学院，北京 102206；
    2. 中国科学院计算技术研究所前瞻研究实验室，北京 100190；
    3. 太仓中科信息技术研究院，江苏 太仓 215400
  • 出版日期:2022-06-30 发布日期:2022-06-28
  • 基金资助:
    国家自然科学基金项目(61972379)

Product surface defect detection and segmentation based on anomaly detection

  1. 1. School of Control and Computer Engineering, North China Electric Power University, Beijing 102206, China;
    2. Prospective Research Laboratory, Institute of Computing Technology, Chinese Academy of Sciences, Beijing 100190, China;
    3. Taicang Institute of Information Technology, Taicang Jiangsu 215400, China
  • Online:2022-06-30 Published:2022-06-28
  • Supported by:
    National Natural Science Foundation of China (61972379)

PDF (PC)

163

摘要：

工业制造中缺陷样本难以获得且缺陷表现形式多样，只用训练正样本的异常检测技术越来越多地被应用于产品表面缺陷检测。异常检测一般通过评估产品图像的异常分数对产品进行有无缺陷的判断，缺乏对缺陷位置的描述，最新提出的异常分割方法对此进行了改进，但对缺陷区域的分割不够精确。基于异常检测方法，使用标准化流来判断产品表面是否有缺陷，采用多尺度特征融合并对齐来初步定位缺陷位置，结合梯度和最大信息熵，使用分水岭算法对初定位结果进行优化得到缺陷分割掩码。在丽盛制板，KolektorSDD 和 AITEX3 个表面缺陷数据集的检测与分割结果均优于其他同类方法。此外，在小样本数据集上也能达到良好的检测与分割精度。

关键词: 异常检测, 缺陷分割, 多尺度特征融合, 特征对齐, 分水岭算法

Abstract:

In industrial manufacturing, it is difficult to obtain defective samples and the defects are in diverse forms. Anomaly detection, which only trains positive samples, is being increasingly applied to defect detection on product surfaces. Anomaly detection generally determines whether the product has defects by evaluating the anomaly score of the product image, while unable to describe the locations of the defects. The latest anomaly segmentation method has been improved, but the segmentation of the defective area is not accurate enough. Based on the anomaly detection method, normalization flow was employed to judge whether the product surface was defective, and multi-scale feature fusion and alignment were adopted to initially locate the defects. Combined with the gradient and maximum information entropy, the watershed algorithm was used to optimize the initial positioning results to obtain the defect segmentation mask. The detection and segmentation results on the three surface defect datasets of Lisheng Board, KolektorSDD, and AITEX are superior to other similar methods. In addition, good detection and segmentation accuracy can also be achieved on few-shots.

Key words: anomaly detection, defect segmentation, multi-scale feature fusion, feature alignment, watershed algorithm

中图分类号: