EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Reference Mixed Semi Supervised Segmentation of Warehouse Defect Images Under Severe Label Scarcity and Sparse Annotations Naomi Reed, Patrick Coleman

Naomi Reed, Patrick Coleman

Journal of the Artificial Intelligence Engineers Consortium · 2026

DOI: 10.66238/jaiec61 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of the Artificial Intelligence Engineers Consortium, 2026. Verfasst von Naomi Reed und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The rapid expansion of global e-commerce has placed unprecedented operational demands on modern logistics centers, necessitating rigorous structural health monitoring of warehouse environments. Automated visual inspection systems driven by deep learning offer a scalable solution for detecting structural defects such as concrete spalling, pallet rack deformation, and surface oxidation. However, training robust semantic segmentation models for these specific industrial applications is severely bottlenecked by the exorbitant cost and domain expertise required for pixel-level annotation. To address this challenge of extreme label scarcity, we propose a novel Reference-Mixed Semi-Supervised Segmentation framework specifically tailored for warehouse defect imagery. Our approach leverages a dual-branch teacher-student architecture coupled with a unique structural augmentation strategy. By dynamically mixing unlabeled target images with highly curated, unannotated reference images depicting known structural states, the model forces the representation space to learn invariant defect features irrespective of background clutter and varying illumination. Furthermore, we introduce a textual morphology-based refinement protocol to filter low-confidence pseudo-labels during the consistency regularization phase. Comprehensive evaluations on a custom warehouse infrastructure dataset demonstrate that our proposed framework achieves competitive segmentation performance utilizing only a marginal fraction of annotated data, significantly outperforming conventional semi-supervised baselines. The methodology not only alleviates the annotation burden but also enhances the generalization capabilities of defect detection models across diverse logistical environments.

Zitieren

APA

Reed, N., & Coleman, P. (2026). Reference Mixed Semi Supervised Segmentation of Warehouse Defect Images Under Severe Label Scarcity and Sparse Annotations Naomi Reed, Patrick Coleman. Journal of the Artificial Intelligence Engineers Consortium. https://doi.org/10.66238/jaiec61

Vancouver

Reed N, Coleman P. Reference Mixed Semi Supervised Segmentation of Warehouse Defect Images Under Severe Label Scarcity and Sparse Annotations Naomi Reed, Patrick Coleman. Journal of the Artificial Intelligence Engineers Consortium. 2026. doi:10.66238/jaiec61

BibTeX
@article{Reed2026,
  author = {Reed, Naomi and Coleman, Patrick},
  title = {Reference Mixed Semi Supervised Segmentation of Warehouse Defect Images Under Severe Label Scarcity and Sparse Annotations Naomi Reed, Patrick Coleman},
  journal = {Journal of the Artificial Intelligence Engineers Consortium},
  year = {2026},
  doi = {10.66238/jaiec61}
}
Kostenlos starten →