Dual flow reverse distillation for unsupervised anomaly detection

Xueqin Jiang, Kai Huang, Shubo Zhou, Weiyu Hu, Huanchun Peng, Jiangliang Jin, Zhijun Fang

Abstract: Highlights•A novel dual flow reverse distillation model separates information into global context and local detail flows.•ACBE and GCEB modules effectively capture global context for detecting logical anomalies.•MCAE module specializes in local detail reconstruction for structural anomaly detection.•Single student decoder unifies context and detail processing for parameter - efficient anomaly detection.•State-of-the-art performance achieved on BTAD dataset with competitive results on MVTec benchmarks.