Industrial Image Anomaly Detection via Synthetic-Anomaly Contrastive Distillation
Industrial image anomaly detection plays a critical role in intelligent manufacturing by automatically identifying defective products through visual inspection. While unsupervised approaches eliminate dependency on annotated anomaly samples, current teacher–student framework-based methods still face...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-06-01
|
| Series: | Sensors |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1424-8220/25/12/3721 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Industrial image anomaly detection plays a critical role in intelligent manufacturing by automatically identifying defective products through visual inspection. While unsupervised approaches eliminate dependency on annotated anomaly samples, current teacher–student framework-based methods still face two fundamental limitations: insufficient discriminative capability for structural anomalies and suboptimal anomaly feature decoupling efficiency. To address these challenges, we propose a Synthetic-Anomaly Contrastive Distillation (<i>SACD</i>) framework for industrial anomaly detection. <i>SACD</i> comprises two pivotal components: (1) a reverse distillation (RD) paradigm whereby a pre-trained teacher network extracts hierarchically structured representations, subsequently guiding the student network with inverse architectural configuration to establish hierarchical feature alignment; (2) a group of feature calibration (<i>FeaCali</i>) modules designed to refine the student’s outputs by eliminating anomalous feature responses. During training, <i>SACD</i> adopts a dual-branch strategy, where one branch encodes multi-scale features from defect-free images, while a Siamese anomaly branch processes synthetically corrupted counterparts. <i>FeaCali</i> modules are trained to strip out a student’s anomalous patterns in anomaly branches, enhancing the student network’s exclusive modeling of normal patterns. We construct a dual-objective optimization integrating cross-model distillation loss and intra-model contrastive loss to train <i>SACD</i> for feature alignment and discrepancy amplification. At the inference stage, pixel-wise anomaly scores are computed through multi-layer feature discrepancies between the teacher’s representations and the student’s refined outputs. Comprehensive evaluations on the MVTec AD and BTAD benchmark demonstrate that our method is effective and superior to current knowledge distillation-based approaches. |
|---|---|
| ISSN: | 1424-8220 |