家用燃气表外观缺陷的改进ViT+FastFlow检测方法研究

Research on the improved ViT+FastFlow detection method for appearance defects of domestic gas meters

外观质量是家用燃气表(DGM)国家强制检定项目之一,针对DGM外观质量检定中匮乏缺陷样本使基于有监督学习检测方法难以泛化到实际应用场景问题,本文研究DGM外观缺陷无监督检测方法,引入Vision Transformer(ViT)改进版EfficientFormerV2-l提取正常样本特征,融合底层和高层特征图,并通过二维标准化流FastFlow将正常特征图映射到标准高斯分布,外观缺陷因离散落在分布以外使异常得分相比正常样本更高,通过设置自适应阈值识别并定位DGM外观缺陷。实验采集DGM正常样本、真实缺陷样本、合成缺陷样本作为数据集并优化检测模型参数,优化后检测模型在图像级别指标AUROC达99.77%,在像素级别指标AUPRO达96.3%,每秒可检测4张以上DGM图像,表明本文方法能准确高效识别与定位DGM外观缺陷。

Appearance quality is one of the national mandatory verification for domestic gas meters(DGM).In view of the lack of defect samples in the appearance quality verification of DGM,which makes the detection method based on supervised learning difficult to generalize to the actual application scenario.This paper studies the unsupervised detection method of DGM appearance defects.EfficientFormerV2-l,the improved Vision Transformer(ViT),is introduced to extract normal sample features,fuse the bottom and highlevel feature maps,and map the normal features to the standard Gaussian distribution using two-dimensional normalizing flow called FastFlow.The appearance defects are scattered outside the distribution so that the abnormal score is higher than the normal sample.By setting an adaptive threshold,the DGM appearance defects are identified and located.The experiment collects DGM normal samples,real defect samples,synthetic defect samples as data sets and optimizes the detection model parameters.The optimized detection model achieves 99.77%AUROC at image level indicators,96.3%AUPRO at pixel level indicators,and can detect more than 4 DGM images per second,indicating that the method in this paper can accurately and efficiently identify and locate DGM appearance defects.