Auxiliary reverse distillation for fabric anomaly detection and localization

Abstract

Fabric anomaly detection plays an important role in ensuring textile quality and improving production efficiency. However, it remains challenging due to the subtle appearance of defects and the limited availability of anomalous samples. Unsupervised anomaly detection methods based on knowledge distillation have demonstrated promising performance, but their effectiveness is often constrained by the optimization difficulty of feature transformation bottlenecks under normal-only training. In this work, we propose an Auxiliary Reverse Distillation (ARD) framework for fabric anomaly detection and localization. The proposed framework introduces an attention-enhanced bottleneck design to improve the trainability and representation capability of the distillation process. By incorporating attention-based feature recalibration and multi-scale feature consistency constraints, ARD facilitates the learning of discriminative normal texture representations and enhances the discrepancy between normal and anomalous regions during inference. Extensive experiments are conducted on the proprietary Color-F fabric dataset, MVTec, DAGM, and several one-class anomaly detection benchmarks. On the Color-F dataset, ARD achieves an average image-level AUROC of 97.4%, outperforming the original reverse distillation method. Furthermore, the proposed framework demonstrates competitive performance on public benchmarks, validating its effectiveness and general applicability for fabric anomaly detection and localization. The results indicate that improving the trainability and representation capability of the distillation bottleneck is an effective strategy for developing accurate and robust textile inspection systems.