This paper considers zero-shot Anomaly Detection (AD), a valuable yet
under-studied task, which performs AD without any reference images of the test
objects. Specifically, we employ a language-guided strategy and propose a
simple-yet-effective architecture CLIP-AD, leveraging the superior zero-shot
classification capabilities of the large vision-language model CLIP. A natural
idea for anomaly segmentation is to directly calculate the similarity between
text/image features, but we observe opposite predictions and irrelevant
highlights in the results. Inspired by the phenomena, we introduce a Staged
Dual-Path model (SDP) that effectively uses features from various levels and
applies architecture and feature surgery to address these issues. Furthermore,
delving beyond surface phenomena, we identify the problem arising from
misalignment of text/image features in the joint embedding space. Thus, we
introduce a fine-tuning strategy by adding linear layers and construct an
extended model SDP+, further enhancing the performance. Abundant experiments
demonstrate the effectiveness of our approach, e.g., on VisA, SDP outperforms
SOTA by +1.0/+1.2 in classification/segmentation F1 scores, while SDP+ achieves
+1.9/+11.7 improvements.

本文介绍了一种利用大型视觉语言模型 CLIP 的零样本异常检测方法 CLIP-AD，采用语言引导策略，通过引入 Staged Dual-Path 模型来解决相似度计算、关键特征、文本与图像特征等问题，进一步引入线性层进行 fine-tuning 来增强性能。实验证明，该方法在分类 / 分割 F1 分数上优于现有方法 1.0/1.2，并且经过扩展的模型 SDP + 可以获得进一步的改进，提高了分类 / 分割 F1 分数 1.9/11.7。