EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Semi-Supervised Contrastive Learning and Entropy Generalization for Rail Surface Defect Segmentation

Bosen Wang, Zelong You, Runliang Tian, Dongyu Fan, Ruifeng Ding, Lei Gao

Journal of Imaging · 14.11.2026 · 12 (11) · 577

DOI: 10.3390/jimaging12110577 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Imaging, online erschienen am 14.11.2026. Verfasst von Bosen Wang und 5 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

To address the challenge of scarce and expensive pixel-level annotations in rail surface defect detection, semi-supervised semantic segmentation offers a practical solution. However, existing semi-supervised frameworks often suffer from insufficient pixel-level feature discrimination and overfitting to limited labeled samples, leading to degraded robustness when dealing with complex defect textures and ambiguous background regions. To address these challenges, we propose SCEG, a Semi-supervised Contrastive Learning with Entropy Generalization framework for rail surface defect segmentation. Specifically, SCEG introduces a Pseudo-label Contrastive Learning (PCL) module that exploits pseudo-labeled unlabeled samples to construct pixel-wise contrastive pairs, thereby enhancing feature discriminability. In addition, an Entropy Distribution Generalization (EDG) strategy is proposed as a complementary regularizer that improves model generalization by regularizing entropy maps across labeled and unlabeled images, where paired pixels may belong to different semantic categories and prediction consistency is therefore not enforced. Experimental validation was conducted on the RSDDs and RSDDs-AUG dataset under various annotation ratios. The experimental results show that the proposed method achieves better mIoU performance than the SOTA comparison methods on both datasets. Under the 1/16 labeled sample condition of RSDDs, an mIoU of 69.35% was achieved, which is 5.5 percentage points (pp) higher than the baseline method. Ablation experiments further confirm the effectiveness of this method in segmenting surface defects on the track.

Zitieren

APA

Wang, B., You, Z., Tian, R., Fan, D., Ding, R., & Gao, L. (2026). Semi-Supervised Contrastive Learning and Entropy Generalization for Rail Surface Defect Segmentation. Journal of Imaging, 12(11), 577. https://doi.org/10.3390/jimaging12110577

Vancouver

Wang B, You Z, Tian R, Fan D, Ding R, Gao L. Semi-Supervised Contrastive Learning and Entropy Generalization for Rail Surface Defect Segmentation. Journal of Imaging. 2026;12(11):577. doi:10.3390/jimaging12110577

BibTeX
@article{Wang2026,
  author = {Wang, Bosen and You, Zelong and Tian, Runliang and Fan, Dongyu and Ding, Ruifeng and Gao, Lei},
  title = {Semi-Supervised Contrastive Learning and Entropy Generalization for Rail Surface Defect Segmentation},
  journal = {Journal of Imaging},
  year = {2026},
  volume = {12},
  number = {11},
  pages = {577},
  doi = {10.3390/jimaging12110577}
}
Kostenlos starten →