One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection

TL;DR

The paper tackles the challenge of robust few-shot anomaly detection by moving beyond direct query-reference feature comparisons. It introduces anomaly personalization via a personalized, anomaly-free diffusion model to transform queries toward the normal manifold, and a triplet contrastive anomaly inference that fuses personalized, anomaly-free, and text-prompt signals. Across 11 datasets in industrial, medical, and semantic domains, the approach achieves consistent improvements over state-of-the-art methods and can augment existing AD techniques with generated anomaly-free samples. This work offers practical gains in accuracy and stability, with potential wide impact in industrial inspection and medical anomaly screening, while noting limitations in zero-shot generalization and proposing future work on open-vocabulary and real-time deployment.

Abstract

Traditional Anomaly Detection (AD) methods have predominantly relied on unsupervised learning from extensive normal data. Recent AD methods have evolved with the advent of large pre-trained vision-language models, enhancing few-shot anomaly detection capabilities. However, these latest AD methods still exhibit limitations in accuracy improvement. One contributing factor is their direct comparison of a query image's features with those of few-shot normal images. This direct comparison often leads to a loss of precision and complicates the extension of these techniques to more complex domains--an area that remains underexplored in a more refined and comprehensive manner. To address these limitations, we introduce the anomaly personalization method, which performs a personalized one-to-normal transformation of query images using an anomaly-free customized generation model, ensuring close alignment with the normal manifold. Moreover, to further enhance the stability and robustness of prediction results, we propose a triplet contrastive anomaly inference strategy, which incorporates a comprehensive comparison between the query and generated anomaly-free data pool and prompt information. Extensive evaluations across eleven datasets in three domains demonstrate our model's effectiveness compared to the latest AD methods. Additionally, our method has been proven to transfer flexibly to other AD methods, with the generated image data effectively improving the performance of other AD methods.

Figures (16)

• Figure 1: Overview of our proposed anomaly personalization approach. First, we use few-shot normal images to customize an anomaly-free diffusion model (top left), similar to Dreambooth. Next, we perform one-to-normal personalization of the query image, obtaining the personalized image (bottom left). Finally, we adopt a triplet contrastive anomaly inference method, which integrates anomaly scores from three aspects to obtain the final results. $S_N$, $S_P$ and $S_{text}$ represent different contrastive anomaly inference processes.
• Figure 2: Visualization of representative results for pixel-level anomaly localization of our proposed method on different datasets.
• Figure 3: Visualizations of anomaly personalization. The red box of the query image indicates anomalous regions.
• Figure 4: The effectiveness of our generated anomaly-free samples for other AD methods. The red numbers highlight performance increases.
• Figure 5: Lists of state and template level prompts employed in this paper to construct text features.