Department of Computer Engineering

In industrial environments, visual inspection done by workers is time-consuming and prone to errors. To overcome these problems, deep learning methods were proposed as a promising solution. To be applicable in real-world systems, these methods should achieve high precision while also being compatible with edge devices, having low computational cost, and supporting real-time use with short inference time. In this study, an unsupervised anomaly detection method using one-class classification was presented. The model uses multi-level features and is trained with Deep SVDD and a regularization loss. The model was tested using MVTec AD and VISA datasets. In addition, the AUROC, AP, and inference time were compared with four unsupervised feature embedding methods. The results demonstrate that the proposed method ran faster on both datasets and achieved acceptable AUROC and AP values (higher than 90%). Future work will focus on improving accuracy while keeping the model efficient.