Unsupervised Anomaly Detection and Segmentation on Dirty Datasets

Future Internet

Lu Wang¹  Xiaohuo Yu¹  Jiahao Guo¹ 

[1] School of Computer Engineering and Science, Shanghai University, Shanghai 200444, China;

关键词: anomaly detection and segmentation;    unsupervised method;    noise resistance;   

DOI  :  10.3390/fi14030086

来源: DOAJ

【 摘 要 】

Industrial quality control is an important task. Most of the existing vision-based unsupervised industrial anomaly detection and segmentation methods require that the training set only consists of normal samples, which is difficult to ensure in practice. This paper proposes an unsupervised framework to solve the industrial anomaly detection and segmentation problem when the training set contains anomaly samples. Our framework uses a model pretrained on ImageNet as a feature extractor to extract patch-level features. After that, we propose a trimming method to estimate a robust Gaussian distribution based on the patch features at each position. Then, with an iterative filtering process, we can iteratively filter out the anomaly samples in the training set and re-estimate the Gaussian distribution at each position. In the prediction phase, the Mahalanobis distance between a patch feature vector and the center of the Gaussian distribution at the corresponding position is used as the anomaly score of this patch. The subsequent anomaly region segmentation is performed based on the patch anomaly score. We tested the proposed method on three datasets containing the anomaly samples and obtained state-of-the-art performance.

【 授权许可】

Unknown