EUVIMEDEUVIMED entdecken
Crossref

HPL-CLIP: hyperbolic prompt learning for zero-shot anomaly detection

Mingjv Peng, Xin Song, Qiyue Zuo, Siyang Xu, Yang Yue

Measurement Science and Technology · 17.07.2026 · 37 (28) · 286115

DOI: 10.1088/1361-6501/ae8713 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 16.07.2026. Verfasst von Mingjv Peng und 4 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Gefördert von: Open Project Program of Marine Ecological Restoration and Smart Ocean Engineering Research Center of Hebei Province, Science Research Project of Hebei Education Department, Natural Science Foundation of Hebei Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Zero-shot anomaly detection (ZSAD) aims to identify anomalies without requiring any target-domain data, making it important for ensuring product quality in industrial manufacturing. Recently, Contrastive Vision-Language Pretraining models have attracted widespread attention in ZSAD. However, existing single-level text prompts need to trade off between classification and segmentation, which abnormal characteristics at different levels are difficult to capture. Meanwhile, general contrastive language-image pre-training (CLIP) models focus on global visual semantics, whereas the anomaly detection task requires fine-grained local visual semantics. To address these issues, we propose a novel ZSAD framework to enhance the semantic discriminability, namely hyperbolic prompt learning (HPL)-CLIP, which fine-tunes text prompts to match multi-level visual features in hyperbolic space. Specifically, we first propose the hyperbolic prompt learning mechanism to learn generic states of normality and abnormality. These generic states are aligned with global and local visual features to capture anomalies. Then, the adaptive pixel-alignment adapter module is designed to extract fine-grained local image details. Finally, we derive a hierarchical enhanced hinge loss to sharpen the boundary between normal and abnormal pixels while encouraging a more compact score distribution for pixels under the same state. Experiments indicate that HPL-CLIP accurately detects anomalous images and localizes anomalous regions. For example, HPL-CLIP outperforms AnomalyCLIP on MVTec AD by 2.1%/0.8% at the image-level AUROC/AP and 0.3%/4.9% at the pixel-level AUROC/PRO, respectively.

Förderung

  • Open Project Program of Marine Ecological Restoration and Smart Ocean Engineering Research Center of Hebei Province
  • Science Research Project of Hebei Education Department
  • Natural Science Foundation of Hebei Province

Zitieren

APA

Peng, M., Song, X., Zuo, Q., Xu, S., & Yue, Y. (2026). HPL-CLIP: hyperbolic prompt learning for zero-shot anomaly detection. Measurement Science and Technology, 37(28), 286115. https://doi.org/10.1088/1361-6501/ae8713

Vancouver

Peng M, Song X, Zuo Q, Xu S, Yue Y. HPL-CLIP: hyperbolic prompt learning for zero-shot anomaly detection. Measurement Science and Technology. 2026;37(28):286115. doi:10.1088/1361-6501/ae8713

BibTeX
@article{Peng2026,
  author = {Peng, Mingjv and Song, Xin and Zuo, Qiyue and Xu, Siyang and Yue, Yang},
  title = {HPL-CLIP: hyperbolic prompt learning for zero-shot anomaly detection},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {28},
  pages = {286115},
  doi = {10.1088/1361-6501/ae8713}
}
Kostenlos starten →