MvP-Diff: Multivariate yet precise diffusion for anomaly images synthesis and segmentation

This paper delves into the simultaneous synthesis of anomaly image samples and their corresponding segmentation labels, addressing the challenge of limited anomaly training data. Existing approaches typically rely on fixed, scarce masks from limited datasets to guide anomaly generation, which constrains diversity in synthesized anomalous regions. Moreover, directly using these guidance masks as labels often results in mask drift issues. To overcome these limitations, we present the Multivariate yet Precise Diffusion model (MvP-Diff). Our method enhances diversity through randomly generated bounding boxes that guide the creation of anomalous regions with varied sizes, shapes, and textures. To resolve mask drift, we introduce an anomaly segmentation label extraction module that computes pixel-level differences between synthesized anomaly images and normal images to generate accurate mask labels. Extensive experiments validate that MvP-Diff produces realistic and diverse anomaly samples while significantly boosting downstream anomaly segmentation performance, demonstrating consistent IoU improvements of 1.47% on PRN and 2.74% on DevNet.