自监督结构化重建的工业图像异常检测与定位  

Structural reconstruction of industrial image anomaly detection and localization based on self supervised learning

作　　者：范伟平 程凡永 张明艳 FAN Weiping;CHENG Fanyong;ZHANG Mingyan(School of Electrical Engineering,Anhui University of Engineering,Wuhu 241000,China)

机构地区：[1]安徽工程大学电气工程学院,安徽芜湖241000

出　　处：《哈尔滨商业大学学报(自然科学版)》2025年第2期152-160,共9页Journal of Harbin University of Commerce:Natural Sciences Edition

基　　金：国家自然科学基金面项目(61976005);芜湖市重点研发项目(2022yf42);检测技术与节能装置安徽省重点实验室开放研究基金(JCKJ2021B06);安徽工程大学-鸠江区产业协同创新项目(2022cyxtb10)。

摘　　要：针对工业图像表面异常检测中缺陷训练样本不足、漏检率高等问题,设计了一种基于自监督结构化重建的工业图像异常检测与定位模型,该模型仅采用正常样本训练即可准确检测和定位异常区域.模型基于自编码器重建输入图像,由于其泛化性能强,异常区域也会被很好地重构,输入图像和重构图像之间的异常区域重构误差减小,从而降低异常检测精度.提出了多互补掩码融合的方法降低重建异常区域可能性,同时保持正常区域能较好地重建,提高重构误差识别异常的准确性.通过评估输入图像和融合重构图像之间的多尺度结构相似性损失实现异常检测和定位.实验结果表明,在图像级和像素级的AUC标准度量下,该方法在MVTec AD数据集上比其他两种先进方法异常检测率分别提高了1.6%和3.4%;异常定位率分别提高了1.7%和4.5%,具有更有效的检测性能.To solve the issues of insufficient defect training samples and high missing rates in the surface anomaly detection of industrial images,a self-supervised structural reconstruction-based anomaly detection and localization model was designed.The model could accurately detect and locate abnormal regions using only normal sample training.The model reconstructed the input image based on an autoencoder.Due to its strong generalization performance,the abnormal region was also well reconstructed,leading to a reduction in thereconstruction error of the abnormal region between the input image and the reconstructed image,and thus decreasing the anomaly detection accuracy.Themulti-complementary mask fusion method was proposed to reduce the likelihood of reconstructing the abnormal region while ensuring that the normal region was well reconstructed,thereby improving the accuracy of anomaly reconstruction error identification.Anomaly detection and localization were achieved by evaluating the multi-scale structural similarity loss between the input image and the fused reconstructed image.Experimental results showed that the anomaly detection rate of the proposed method was 1.6%and 3.4%higher than that of the other two advanced methods on the MVTec AD dataset under the image-level and pixel-level AUC standards,respectively.The abnormal localization rate was increased by 1.7%and 4.5%,respectively,demonstrating more effective detection performance.

关 键 词：自监督 异常检测与定位 结构化重建 掩码重构 融合算法 结构相似性损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]