EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

KKA: Knowledge Augmentation From Large Language Models for Category‐Level Vision Anomaly Detection

Dong Chen, Zhengqing Hu, Liuyuan Su, Peiguang Fan, Yueting Zhuang, Hua Wang, Xiaoheng Jiang, Mingliang Xu

CAAI Transactions on Intelligence Technology · 13.11.2026

DOI: 10.1049/cit2.70196 ↗

Kurz gefasst

Fachartikel in der Zeitschrift CAAI Transactions on Intelligence Technology, online erschienen am 13.11.2026. Verfasst von Dong Chen und 7 weiteren Autor:innen. Die Arbeit nennt 50 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Natural Science Foundation of China, China Postdoctoral Science Foundation.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Category‐level vision anomaly detection, particularly in unsupervised settings, often struggles to distinguish between normal samples and anomalies due to the wide variability in anomalies. Recently, an increasing number of studies have focused on generating anomalies to help detectors learn more effective boundaries between normal samples and anomalies. However, as the generated anomalies are often derived from random factors, they frequently lack realism and offer limited support in constructing effective boundaries. To address these challenges, we propose Key Knowledge Augmentation (KKA), which leverages anomaly‐related knowledge extracted from Large Language Models (LLMs) to improve the recognition of category‐level semantic anomalies. More specifically, KKA leverages the extensive prior knowledge of LLMs to generate meaningful anomalies based on normal samples. Then, KKA classifies the generated anomalies as easy anomalies and hard anomalies according to their similarity to normal samples. Easy anomalies exhibit significant differences from normal samples, whereas hard anomalies closely resemble normal samples. KKA iteratively refines the generated anomalies by progressively increasing the proportion of hard anomalies, enabling lightweight detectors to learn more discriminative decision boundaries while maintaining substantially lower inference‐time overhead than LLM‐based detectors. Experimental results show that the proposed method significantly improves the performance of various category‐level vision anomaly detectors. The code for KKA can be found at https://github.com/Anfeather/KKA .

Förderung

  • National Natural Science Foundation of China
  • China Postdoctoral Science Foundation

Zitieren

APA

Chen, D., Hu, Z., Su, L., Fan, P., Zhuang, Y., Wang, H., Jiang, X., & Xu, M. (2026). KKA: Knowledge Augmentation From Large Language Models for Category‐Level Vision Anomaly Detection. CAAI Transactions on Intelligence Technology. https://doi.org/10.1049/cit2.70196

Vancouver

Chen D, Hu Z, Su L, Fan P, Zhuang Y, Wang H, et al. KKA: Knowledge Augmentation From Large Language Models for Category‐Level Vision Anomaly Detection. CAAI Transactions on Intelligence Technology. 2026. doi:10.1049/cit2.70196

BibTeX
@article{Chen2026,
  author = {Chen, Dong and Hu, Zhengqing and Su, Liuyuan and Fan, Peiguang and Zhuang, Yueting and Wang, Hua and Jiang, Xiaoheng and Xu, Mingliang},
  title = {KKA: Knowledge Augmentation From Large Language Models for Category‐Level Vision Anomaly Detection},
  journal = {CAAI Transactions on Intelligence Technology},
  year = {2026},
  doi = {10.1049/cit2.70196}
}
Kostenlos starten →