Adaptive Loss Weighting for Unsupervised Industrial Anomaly Detection via Reverse Distillation

Abstract

Due to the use of fixed loss weight allocation between different feature layers, unsupervised industrial surface anomaly detection often fails to capture multi-scale anomalies, resulting in the loss of fine-grained defects and severe background noise. To address this research gap, this paper proposes an unsupervised anomaly detection framework with an adaptive layer-weighted loss mechanism. By embedding a learnable softmax modulation parameter module in the reverse distillation network, this method dynamically optimizes the loss contributions of shallow spatial features and deep semantic representations during the training phase. Experiments were conducted on the MVTec AD benchmark dataset. Importantly, the proposed method demonstrates an extremely fast convergence speed, significantly improving localization and classification performance in just ten training epochs, especially for challenging categories such as capsules, zippers, and screws, with image-level AUROC values increased by up to 12.7%. The results show that dynamic loss weighting can effectively suppress edge noise and enhance sensitivity to multi-scale defects, providing an efficient and robust solution for real-time automated industrial quality inspection.