EUVIMEDEUVIMED entdecken
Crossref

Pose-robust and context-aware industrial few-shot anomaly detection via deformable alignment

WenTing He, FengYu Yang, WuYi Li, ShiHua Yu, ZhiJie Chen, Ying Ma

Measurement Science and Technology · 31.07.2026 · 37 (30) · 305002

DOI: 10.1088/1361-6501/ae8d48 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Measurement Science and Technology, online erschienen am 31.07.2026. Verfasst von WenTing He und 5 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Gefördert von: General Program of the Natural Science Foundation of Jiangxi Province, the Key Research and Devel opment Program of Xinjiang Uygur Autonomous Region, the Key Program of the Joint Fund of the Natural Science Foun dation of Jiangxi Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Industrial few-shot anomaly detection (FSAD) is important for automated quality inspection because it must identify defects from only a few normal reference samples. Most existing FSAD methods rely on fixed-grid feature extraction and therefore implicitly assume strict spatial alignment between support and query images. In practical industrial settings, however, pose perturbations such as translation, rotation, and non-rigid deformation often violate this assumption, leading to feature drift and unreliable anomaly scoring. To address this issue, we propose PRCANet, a pose-robust and context-aware FSAD framework. PRCANet consists of a pose-robust deformable alignment module for distance-guided feature calibration, an alignment-quality-guided reliability reweighting mechanism for suppressing unreliable forced alignments, and a Context-Aware Statistical Fusion (CA 3 ) module for stable image-level aggregation. Experiments on the MVTec AD and Manufacturing Pose-Distorted Dataset (MPDD) benchmarks show that PRCANet achieves competitive overall performance and shows particular advantages in pose-challenging scenarios. Under the 2-shot setting on MPDD, it reaches an image-level AUROC of 82.1%, outperforming the baseline by 14.4%, while maintaining an inference speed of 13.91 FPS. These results suggest the practical potential of PRCANet for real-time industrial inspection under geometric variation.

Förderung

  • General Program of the Natural Science Foundation of Jiangxi Province
  • the Key Research and Devel opment Program of Xinjiang Uygur Autonomous Region
  • the Key Program of the Joint Fund of the Natural Science Foun dation of Jiangxi Province

Zitieren

APA

He, W., Yang, F., Li, W., Yu, S., Chen, Z., & Ma, Y. (2026). Pose-robust and context-aware industrial few-shot anomaly detection via deformable alignment. Measurement Science and Technology, 37(30), 305002. https://doi.org/10.1088/1361-6501/ae8d48

Vancouver

He W, Yang F, Li W, Yu S, Chen Z, Ma Y. Pose-robust and context-aware industrial few-shot anomaly detection via deformable alignment. Measurement Science and Technology. 2026;37(30):305002. doi:10.1088/1361-6501/ae8d48

BibTeX
@article{He2026,
  author = {He, WenTing and Yang, FengYu and Li, WuYi and Yu, ShiHua and Chen, ZhiJie and Ma, Ying},
  title = {Pose-robust and context-aware industrial few-shot anomaly detection via deformable alignment},
  journal = {Measurement Science and Technology},
  year = {2026},
  volume = {37},
  number = {30},
  pages = {305002},
  doi = {10.1088/1361-6501/ae8d48}
}
Kostenlos starten →