MSSF: A Multi-scale Siamese Flow Architecture for Multi-texture Class Anomaly Detection
Published in Pattern Recognition: 27th International Conference, ICPR 2024, Kolkata, India, December 1–5, 2024, Proceedings, Part XVII (Lecture Notes in Computer Science, vol 15317), 2024
This paper proposes a lightweight flow-based pipeline called Multi-Scale Siamese Flow (MSSF) for multi-texture class anomaly detection. It integrates Multi-level Feature Fusion (MLFF) to fully utilize shallow and deep features, and incorporates a Mixed Anomalies Synthesis (MAS) method with a novel training loss (combining negative log-likelihood and changeable self-supervised hindering loss) for self-supervised training. Experiments on MVTec-AD, KSDD2, MT, and AITEX datasets show its effectiveness, with inference speed twice that of UniAD and a balanced performance-speed trade-off.
Recommended citation: Chen, Y., Hu, Z., Huang, L., Zhang, J. (2025). "MSSF: A Multi-scale Siamese Flow Architecture for Multi-texture Class Anomaly Detection." <Pattern Recognition (ICPR 2024 Proceedings, Part XVII). Lecture Notes in Computer Science, vol 15317. Springer, Cham. https://doi.org/10.1007/978-3-031-78447-7_3 https://link.springer.com/chapter/10.1007/978-3-031-78447-7_3