A Transfer Learning-Based U-Net Approach for Industrial Anomaly Detection with Limited Samples
Kaiwen Yang, Guijie Zhu, Junyuan Zhao, Decheng Ding, Jiafan Zhuang, Chuliang Wei · 2025
In industrial anomaly detection, the scarcity of defective samples often leads to model overfitting and degraded deployment performance. To address this challenge, this paper proposes a U-Net-based transfer learning strategy: the model is first pre-trained on a large-scale road crack segmentation dataset (Crack500) to learn generic features of surface defects, followed by fine-tuning high-level network parameters on downstream datasets (KolektorSDD, RSDDs) containing only limited industrial defect samples for binary segmentation. Experimental results demonstrate that transfer learning improves the F1-score by 12.75% on the KolektorSDD dataset and 40.89% on the RSDDs dataset compared to training from scratch, with a significant boost in mIoU. The proposed method maintains high segmentation accuracy even under small-sample conditions, offering both theoretical and practical insights for industrial anomaly detection.