Efficient Industrial Anomaly Detection via Knowledge Distillation

The field of industrial anomaly detection (AD) using deep learning has seen significant advancements in recent years. However, many state-of-the-art methods require substantial computational resources and exhibit considerable latency, limiting their practical application in real-world industrial systems that demand real-time processing. Recently, detection methods based on knowledge distillation (KD) have emerged as promising alternatives, showing improvements in both anomaly detection performance and inference runtime efficiency. KD-based detection methods leverage the inherent representation divergence between the teacher-student (T-S) models as crucial evidence for AD. In this study, we continue this line of research and introduce EIAD, which stands for Efficient Industrial Anomaly Detection. EIAD consists of two pairs of teacher-student architectures, each specializing in detecting anomalies in products with either complex texture or intricate shape information. Our proposed EIAD achieves outstanding performance, with an image-level anomaly detection AUROC score of up to 99.3% on the widely recognized industrial AD benchmark, MVTec AD, while maintaining inference times in milliseconds.