Few-Shot Industrial Anomaly Detection via Proxy Task Learning on Normal Samples

Traditional industrial anomaly detection methods rely on extensive samples through unsupervised learning using the one-product-one-model paradigm. This data-hungry and rigid training-inference paradigm is not prone to adapting to flexible and changeable real industrial scenarios. Few-shot learning and contrastive learning offer promising solutions by enabling models to recognize anomalies with a few normal samples. This paper presents a novel approach that enhances model robustness and accuracy by leveraging three different proxy tasks learned from normal samples. We generate positive samples through image geometric transformations and negative samples by injecting anomaly noise into normal samples. A multi-level, multi-scale feature fusion structure is proposed to produce refined, fine-grained image feature maps. Specifically, we employ a rotation prediction task, a pixel-level contrastive learning task, and an image-level contrastive learning task to jointly optimize our model. Comparative and comprehensive experiments were conducted on two sought-after industrial anomaly detection datasets. The results demonstrate that our method achieves competitive performance compared to other advanced methods.