基于风格记忆的跨域小样本异常检测
Style-aware Cross Domain Few-Shot Anomaly Detection
-
摘要: 小样本异常检测仅利用少量正常样本进行检测和异常定位. 现有的小样本异常检测方法训练和测试都在相同的域中进行, 对于跨域场景的研究尚处于空白. 因此, 在本文中, 我们关注一个具有挑战性但实用的跨域小样本异常检测问题, 其训练数据和测试数据来自不同的分布. 问题难点在于如何将源域学到的知识迁移到目标域中进行小样本异常检测, 为此我们提出了两个模块. 首先, 我们将域差异定义为风格特征差异, 并设计风格记忆模块存储来自源域的域内统计特征, 并将学习到的特征集成到目标域中, 从而减轻域差异带来的影响, 提升性能. 然后我们引入了对比学习策略用来区分不同类别的增强查询特征, 以进一步加强模型对源域中不同类别的表征能力, 提高小样本检测性能. 最后, 在MVTec AD、MPDD和MVTec LOCO-AD数据集间的跨域小样本异常检测任务上, 与最新小样本方法相比我们的模型都取得了最佳性能, 其中图像AUROC指标提升了3%到5%. 同时在本域内小样本异常检测任务上所提模型仍然有优异的表现.Abstract: Few-shot anomaly detection involves using a limited number of normal samples to detect and localize anomalies. While existing methods for few-shot anomaly detection are primarily applied within the same domain, research on cross-domain scenarios remains limited. In this paper, we address the challenging yet practical problem of cross-domain few-shot anomaly detection, where training and test data are drawn from different distributions. A key challenge in this context is transferring knowledge learned in the source domain to the target domain for effective few-shot anomaly detection. To address this challenge, we propose a two-module approach. The first module defines domain differences in terms of differences in style features and employs a style memory module to store intra-domain statistical features from the source domain. These features are then integrated into the target domain to mitigate the impact of domain differences and enhance performance. The second module introduces a contrastive learning strategy to differentiate between enhanced query features for different categories, thereby improving the model’s ability to represent different categories in the source domain and enhancing few-shot detection performance. Our approach achieves state-of-the-art performance on cross-domain few-shot anomaly detection setting involving MVTec AD, MPDD, and MVTec LOCO-AD datasets, with image AUROC improvements ranging from 3% to 5% compared to baseline methods. Moreover, the proposed model still has the best performance on the few-shot anomaly detection task under the intra-domain setting.