Effective industrial anomaly localization using class-conditioned irredundant feature patch selection

  • Moon, A-Seong; 
  • Jung, Se-Hoon; 
  • Zulkifley, Mohd Asyraf; 
  • Lee, Jaesung
Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

Patch-based anomaly localization compares test patches against a memory bank of normal embeddings, but strongly repetitive textures saturate the bank with redundant patches, thereby shrinking nearest-neighbor margins and increasing background-like false positives. This study introduces a class-conditioned irredundant feature patch selection procedure that suppresses near-duplicate candidates within each class, maximizes the dispersion of selected representatives under a fixed memory budget, and leaves the backbone and scoring rule unchanged. On a standard industrial visual anomaly detection benchmark, the resulting memory yielded consistently higher pixel-level average precision and F1-Score than strong retrieval-based baselines, with particularly large gains on highly repetitive textures and reduced threshold sensitivity during operating-point selection. The source code is available at

키워드

Anomaly localization; Memory bank; Patch embeddings; Irredundant selection; Nearest-neighbor scoring
제목
Effective industrial anomaly localization using class-conditioned irredundant feature patch selection
저자
Moon, A-Seong; Jung, Se-Hoon; Zulkifley, Mohd Asyraf; Lee, Jaesung
DOI
10.1007/s10044-026-01762-4
발행일
2026-08
유형
Article
저널명
Pattern Analysis and Applications
권
29
호
3