Scholarlink

<p>Industrial anomaly detection is a critical component of industrial production processes. Existing deep learning-based approaches have overcome numerous challenges and achieved superior detection outcomes. However, prior studies face significant difficulties in generalizing to unseen object categories during training for real-world industrial applications. To bridge this gap, we propose SupAD, a CLIP-based zero-shot anomaly detection(ZSAD) method. We design a Dynamic Window Partitioner(DWP) embedded within the visual encoder to adjust window sizes based on input image surface complexity dynamically, coupled with a joint texture aware attention mechanism to capture fine-grained anomalous features. Furthermore, we propose a Gated-Agnostic Multimodal Adaptive(GAMA) prompt learning module integrated with a spatial gradient adapter. Finally, Hierarchical Contrastive Learning(HCL) is incorporated to refine both global and local prompt learning, effectively enhancing the model’s cross-domain generalization capabilities. Experimental results demonstrate that our model surpasses existing ZSAD methods on both one private dataset and five public benchmarks, achieving state-of-the-art performance. More importantly, this approach offers significant inspiration for other CLIP-based anomaly detection methodologies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Supad: a superordinary zero-shot industrial anomaly detection network based on gated-agnostic multimodal adaptive learning prompts

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Xinying Li,
  • Junfeng Jing,
  • Tong Wu,
  • Xin Zhang,
  • Wei Liu

摘要

Industrial anomaly detection is a critical component of industrial production processes. Existing deep learning-based approaches have overcome numerous challenges and achieved superior detection outcomes. However, prior studies face significant difficulties in generalizing to unseen object categories during training for real-world industrial applications. To bridge this gap, we propose SupAD, a CLIP-based zero-shot anomaly detection(ZSAD) method. We design a Dynamic Window Partitioner(DWP) embedded within the visual encoder to adjust window sizes based on input image surface complexity dynamically, coupled with a joint texture aware attention mechanism to capture fine-grained anomalous features. Furthermore, we propose a Gated-Agnostic Multimodal Adaptive(GAMA) prompt learning module integrated with a spatial gradient adapter. Finally, Hierarchical Contrastive Learning(HCL) is incorporated to refine both global and local prompt learning, effectively enhancing the model’s cross-domain generalization capabilities. Experimental results demonstrate that our model surpasses existing ZSAD methods on both one private dataset and five public benchmarks, achieving state-of-the-art performance. More importantly, this approach offers significant inspiration for other CLIP-based anomaly detection methodologies.