基于泊松图像融合的自监督缺陷检测方法
Self-supervised Defect Detection Method Based on Poisson Image Fusion
陈腾飞 1戴元杰 2廖杜杰 1朱志鹏 1吴健辉2
作者信息
- 1. 三维重建与智能应用技术湖南省工程研究中心, 湖南 岳阳 414006;湖南理工学院 机械科学与工程学院, 湖南 岳阳 414006
- 2. 三维重建与智能应用技术湖南省工程研究中心, 湖南 岳阳 414006;湖南理工学院 信息科学与工程学院, 湖南 岳阳 414006
- 折叠
摘要
提出一种基于泊松图像融合的自监督缺陷检测方法,采用泊松图像融合对无标注的正常样本进行数据增强,生成多样化的、更贴近实际的模拟缺陷样本,解决缺陷样本数量少且不易标注的问题.结合缺陷样本的特征提出一种CANet网络,引入卷积注意力模块对编码器—解码器结构进行优化,防止采样过程中的信息丢失,并在网络末端添加掩码卷积层以提高输入数据的重建精度.在MV Tec数据集上进行实验,总体检测AUROC达到96.1%;通过与三种典型检测方法的比较,证明所提方法的有效性且具备较好的泛化性,能满足工业生产中不同种类产品的表面缺陷检测要求.
Abstract
In this paper,a self-supervised defect detection method based on Poisson image fusion is investigated,Poisson image fusion is used to augment the data of unlabeled normal samples to generate diversified and more realistic simulated defect samples.This method solves the problem of the small number of defect samples and not easy to label.A CANet network is proposed by combining the characteristics of defective samples,and a convolutional attention module is introduced to optimize the encoder-decoder structure,in order to prevent information loss in the sampling process.At the same time,a masked convolutional layer is added to improve the reconstruction accuracy of the input data at the end of the network.The experimental results on the MV Tec dataset achieved an overall detection AUROC of 96.1%,and the comparison with three typical detection methods further proves the effectiveness of our method with better generalization,which can meet the requirements of surface defect detection for different kinds of products in industrial production.
关键词
泊松图像融合/自监督学习/注意力机制/掩码卷积层Key words
Poisson image fusion/self supervised learning/attention mechanism/masked convolutional layer引用本文复制引用
基金项目
湖南省研究生科研创新项目(CX20221237)
湖南省研究生科研创新项目(CX20221219)
出版年
2024
NETL
NSTL
维普
万方数据