Abstract
Abstract
Anomaly detection is important for finding production faults and plays a key role in efficiency, reliability, and process quality. Since manual visual inspection is time-consuming and prone to errors in industrial environments, deep learning methods were proposed as a promising solution. Many anomaly detection models achieve high performance, but they are impractical to deploy on resource-constrained devices.
To be applicable in real-world systems, anomaly detection methods should be capable of high precision, low computational cost device compatible, and fast to support real-time operation. To bridge the gap between performance and applicability, an unsupervised anomaly detection method, called ALight-IAD, using one-class classification was presented in this study. Multi-level features are extracted from a pretrained network and fused with a channel attention mechanism, and then projected into the embedding space. The model is trained by optimizing via Deep SVDD loss with an additional regularization term. We evaluated the method on widely-used MVTec AD, VISA, BTAD, and MPDD datasets. Additionally, AUROC, inference time, memory consumption, and CO2 equivalent emission were compared against four unsupervised state-of-the-art feature embedding methods, which were retrained for detailed benchmarking in this study. The results demonstrate that the proposed method ran up to 17× faster, achieved comparable performance more than 90% AUROC (MVTec AD dataset), has a lightweight architecture including 6 million parameters, and consumes less energy than compared SOTA methods.