Logical Anomaly Detection in Industrial Images Based on Multi-Combinatorial Knowledge Distillation
Yunfeng Zhu, Qishen Li, Hua Huang, Qiufeng Li · 2025
In the domain of industrial image anomaly detection, challenges such as the paucity of anomaly samples, the heterogeneity of anomaly samples, and the difficulty of detecting fine-grained logical anomalies, a Lightweight Distillation Network (LDN) based on multi-combined knowledge distillation is proposed. The model employs a multi-combination knowledge distillation framework, denoted as EfficientAD-S, as the basic model. The LDN is designed with a depth-separable residual structure (DS-Res2Net) to form a Mixed Convolutional Module (MCM). It is utilized to construct an Autoencoder network architecture, which leverages the improved Autoencoder to reconstruct deep features and to learn the normal characteristics of images. Subsequently, a Feature Extraction Module (FEM) is introduced to further enhance anomaly detection performance while maintaining the lightweight nature of the network. Experimental results show that on the MVTec LOCO AD dataset, the proposed method bring an improvement of 1.3% AUROC compared to the benchmark model, with an anomaly detection accuracy reaching 90.1%.