Dual flow reverse distillation for unsupervised anomaly detection

Abstract

Unsupervised image anomaly detection is crucial in industrial manufacturing due to the difficulty of collecting a diverse set of anomaly samples. Recently, reverse distillation-based methods, with a teacher encoder guiding a student decoder, have shown promising performance. However, existing methods generally focus on identifying only one type of anomaly, either structural anomalies or logical anomalies, and struggle to address both simultaneously. In this paper, we propose a novel dual flow reverse distillation model for anomaly detection, which separates the information flow into global context and local detail sub-flows. The global context sub-flow implemented by the Convolution and Self-Attention Integrated Bottleneck Embedding (ACBE) and the Global Context Embedding Block (GCEB), targets logical anomalies, while the local detail sub-flow implemented by the Multiscale Channel Autoencoder (MCAE), focuses on structural anomalies. Different decoding layers in the student network are then specifically designed to process these information flows, enabling the model to effectively address both types of anomalies. Extensive experiments validate the effectiveness of our method, demonstrating competitive performance on the MVTec and MVTec LOCO datasets, and achieving state-of-the-art results on the more challenging BTAD dataset.