MSSF: A Multi-scale Siamese Flow Architecture for Multi-texture Class Anomaly Detection

This paper addresses the limitation of most multi-class anomaly detection methods that focus on increasing backbone parameters or network depth. Taking multi-texture anomaly detection as an example, it validates the lightweight flow-based pipeline MSSF (Multi-Scale Siamese Flow):

1. It adopts Multi-level Feature Fusion (MLFF) to fully use extracted shallow and deep features, avoiding information loss from single-scale feature processing.
2. It incorporates a Mixed Anomalies Synthesis (MAS) method and trains the pipeline in a self-supervised manner by designing a novel training loss that combines negative log-likelihood with a changeable self-supervised hindering loss.

Extensive experiments on real-world texture subsets or datasets (MVTec-AD, KSDD2, MT, AITEX) confirm the effectiveness of MSSF. Its inference speed surpasses the second fastest method (UniAD) by about 2 times, and it achieves an effective balance between performance and speed compared with other cutting-edge methods. This work provides a lightweight solution for multi-texture class anomaly detection, applicable to scenarios like industrial quality inspection.

Download paper here

Recommended citation: Chen, Y., Hu, Z., Huang, L., Zhang, J. (2025). “MSSF: A Multi-scale Siamese Flow Architecture for Multi-texture Class Anomaly Detection.” «i>Pattern Recognition (ICPR 2024 Proceedings, Part XVII)</</i>. Lecture Notes in Computer Science, vol 15317. Springer, Cham. https://doi.org/10.1007/978-3-031-78447-7_3