TED:Text-Axis Evidence Decomposition for Prompted Anomaly Localization

📄 arXiv: 2609.39033v1 📥 PDF

作者: JinYoung Kim, Geonho Kim, GiJeong Park, Geonu Lee, YoungJoon Yoo

分类: cs.CV, cs.AI

发布日期: 2026-09-30

备注: 40th Conference on Neural Information Processing Systems (NeurIPS 2026)

💡 一句话要点

提出TED以解决CLIP模型在缺陷定位中的不可靠性问题

🎯 匹配领域: 支柱九：具身大模型 (Embodied Foundation Models)

关键词: 异常检测 视觉-语言模型 缺陷定位 后处理评分 多模态学习

📋 核心要点

  1. 现有的CLIP基础异常检测方法在领域转移时，常常对真实缺陷和复杂正常区域赋予高异常分数，导致定位不可靠。
  2. TED通过后处理评分方法，判断模糊响应是由缺陷补丁还是误判的正常补丁支持，从而提高缺陷定位的可靠性。
  3. 实验结果显示，TED在不同的像素级设置下，相较于基线方法P-AUROC、P-PRO和P-AP均有显著提升，尤其在高竞争环境下效果更佳。

📝 摘要（中文）

CLIP是一种强大的视觉-语言模型，但并未针对细粒度缺陷定位进行设计，因此CLIP基础的异常检测器通过提示或轻量级模块进行适应以提高缺陷敏感性。我们发现，增强的敏感性并不一定使局部证据可靠：在领域转移下，适应后的CLIP-AD模型常常对真实缺陷和视觉复杂的正常区域都赋予高异常分数。我们提出TED（文本轴证据分解），一种后处理评分方法，旨在判断每个模糊响应是更好地支持源缺陷补丁还是被误认为异常的源正常补丁。TED在不改变主干和提示的情况下，比较这些支持，并且无需目标领域训练。实验表明，TED在像素级定位上显著提高了性能。

🔬 方法详解

问题定义：论文要解决的问题是现有CLIP基础异常检测模型在领域转移时，无法有效区分真实缺陷与复杂正常区域，导致高异常分数的误判。

核心思路：TED的核心思路是通过后处理评分，分析每个模糊响应的支持来源，判断其是由真实缺陷补丁还是误判的正常补丁支持，从而提高定位的可靠性。

技术框架：TED的整体架构包括两个主要模块：一是对源缺陷补丁和源正常补丁的支持进行比较，二是在主干和提示不变的情况下进行评分。

关键创新：TED的创新点在于其后处理评分机制，能够在不进行目标领域训练的情况下，显著提高像素级定位的准确性，与传统方法相比，TED更能有效解码缺陷证据。

关键设计：TED的设计中，采用了无监督的评分方式，利用预训练的多模态表示进行源校准残差修正，确保在不同竞争环境下的稳定性和可靠性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果表明，TED在不同的像素级设置下，相较于基线方法P-AUROC、P-PRO和P-AP均有显著提升，尤其在高竞争环境下，平均定位增益从低竞争的+5.0提升至中高竞争的约+10.9，显示出TED在复杂场景下的有效性。

🎯 应用场景

该研究的潜在应用领域包括工业缺陷检测、医疗影像分析及安全监控等。通过提高缺陷定位的可靠性，TED能够帮助相关领域的从业者更准确地识别和处理异常情况，进而提升产品质量和安全性。未来，TED的技术也可能扩展到其他视觉-语言任务中，推动多模态学习的发展。

📄 摘要（原文）

CLIP is a powerful vision-language model, but it was not designed for fine-grained defect localization; CLIP-based anomaly detectors therefore adapt it with prompts or lightweight modules to increase defect sensitivity. We show that stronger sensitivity does not necessarily make local evidence reliable: under domain shift, adapted CLIP-AD models often assign high anomaly scores to both true defects and visually complex normal regions. The issue is not simply missing defect information, but a local scoring rule that decodes defect and hard-normal evidence, having the same anomaly evidence. We propose TED (Text-Axis Evidence Decomposition), a post-hoc scoring method that asks whether each ambiguous response is better supported by source defect patches or by source normal patches mistaken as anomalous. TED compares these supports under the host's normal-versus-anomaly text response, leaves the backbone and prompts unchanged, and requires no target-domain training. It works as a train-free score for raw VLM backbones or as a source-calibrated residual correction for adapted CLIP-AD hosts. Across frozen VLM backbones, TED substantially improves pixel-level localization over raw prompt similarity; across adapted hosts, it improves most pixel-level settings over P-AUROC, P-PRO, and P-AP. Gains are largest under stronger hard-FP competition, with mean localization gain increasing from +5.0 in low-competition regimes to about +10.9 in mid/high-competition regimes. These results suggest that recoverable defect evidence can already exist in pretrained multimodal representations, but reliable localization requires decoding it against hard-normal competitors. Code will be released at TED GitHub repository.