CFFDist: Cross-Scale Feature Fusion Distillation Network for Industrial Anomaly Localization

Hui Zhi, Hao Qin, Lanning Zhang, Jie Guo, Bin Song

Abstract: Unsupervised anomaly localization plays a crucial role in detecting surface defects in industrial products, and the knowledge distillation network stands out for its effectiveness in anomaly localization. To further enhance the sensitivity of knowledge distillation networks to anomalies and mitigate the risk of overfitting, while addressing the challenge of student decoder network to accurately reconstruct fine-grained features, we propose an innovative Cross-Scale Feature Fusion Distillation Network (CFFDist). The CFFDist network achieves superior performance through the utilization of a distinctive feature fusion distillation approach and the incorporation of two key modules, namely the cross-scale feature fusion window (CFFW) and the abnormal simulation module (ASM). The CFFW provides a more comprehensive representation of the student network by fusing feature information across neighboring layers. This enables a better understanding of the relationships between characteristics at various levels. Additionally, ASM effectively simulates pseudo anomalies, thereby increasing both the quantity and diversity of anomaly samples. Based on four standard and challenging datasets, experimental results demonstrate that the CFFDist network can handle anomalies with high accuracy. These results provide solid evidence for intelligent detection in actual industrial applications by showcasing the capabilities and strengths of CFFDist.