Search papers, labs, and topics across Lattice.
This paper introduces OPD-IAD, a novel framework that enhances industrial anomaly detection (IAD) by leveraging language judgments to achieve precise pixel-level anomaly localization in large vision-language models (LVLMs). By employing evidence-privileged dense on-policy self-distillation, the method distills defect evidence onto the model's judgment trajectory, allowing for dense supervision and improved anomaly perception. The results demonstrate that OPD-IAD outperforms existing LVLM-based IAD methods across various metrics, establishing a new benchmark in the field.
Language can effectively guide pixel-level anomaly detection without compromising visual fidelity, leading to unprecedented performance in industrial applications.
Large vision-language models (LVLMs) have recently shown strong potential for industrial anomaly detection (IAD) by providing image-level anomaly judgments and interpretable defect reasoning. However, current LVLM-based IAD methods still struggle to produce precise pixel-level anomaly maps from generated language judgments. We aim to achieve precise pixel-level localization while using language as guidance rather than letting it dominate the visual response. Specifically, we propose \textbf{OPD-IAD}, an evidence-privileged dense on-policy self-distillation framework for LVLM-based IAD. OPD-IAD distills privileged defect evidence onto the model's own on-policy judgment trajectory, enabling the final generated judgment to be learned under dense supervision rather than treated only as a textual answer. The resulting judgment serves as a semantic condition for dense anomaly perception. To turn this condition into dense visual evidence, we introduce \textbf{Language-guided Visual Anchoring}, which uses a judgment reforward to re-encode the image and question under the final-judgment condition into semantic anchors and contrasts them with dense visual features through a contrastive heatmap head to generate anomaly maps. The language judgment therefore provides compact semantic guidance, while dense visual features remain the basis for pixel-level scoring, allowing language to guide anomaly localization without letting language quality directly dictate the pixel-level response. Extensive experiments show that OPD-IAD achieves the best overall performance among LVLM-based IAD methods, leading on most image-level, pixel-level, and QA metrics.
