摘要
针对不同异常图像数据间的差异以及深度模型泛化能力不足的问题,提出一种改进掩码自编码器的多类工业图像异常检测方法。首先,利用正常图像样本数据训练改进的掩码自编码器(MAE)模型,使该模型具有正常图像重构的能力;然后,根据改进的MAE模型的重构图像与原图像的差异,区分正常与异常图像数据;最后,利用该方法在公开的工业图像数据集上同时检测多个类别的异常图像数据,平均AUC为0.895,相较于MKD、U-Net、DAGAN,其检测精度分别提高了2.05%、9.28%和2.52%,验证了该方法的有效性。
Aiming at the differences between different abnormal image data and the insufficient generalization ability of deep models,an improved mask autoencoder based multi class industrial image anomaly detection method is proposed.Firstly,an improved Mask Autoencoder(MAE)model is trained using normal image sample data to enable the model to reconstruct normal images;Then,based on the difference between the reconstructed image of the improved MAE model and the original image,distinguish between normal and abnormal image data;Finally,this method was used to simultaneously detect multiple categories of abnormal image data on a publicly available industrial image dataset,with an average AUC of 0.895.Compared with MKD,U-Net,and DAGAN,the detection accuracy was improved by 2.05%,9.28%,and 2.52%,respectively,verifying the effectiveness of this method.
作者
胡洋
肖明
康嘉文
HU Yang;XIAO Ming;KANG Jiawen(Guangdong University of Technology,Guangzhou 510006,China)
出处
《自动化与信息工程》
2024年第4期30-35,共6页
Automation & Information Engineering
基金
广东省科技公关计划(2022A1515011445)
大范围场景空间定位与自然人机交互关键技术(502200027)。
关键词
编解码重构
掩码自编码器
异常检测
工业图像
encoder-decoder reconstruction
masked autoencoder
anomaly detection
industrial image