FDPT-Net (feature focusing diffusion pyramid and task-aligned network): A train wheelset tread detection network based on feature focusing diffusion pyramid and dynamic task alignment
Review of Scientific Instruments · 01.11.2026 · 97 (11)
DOI: 10.1063/5.0336077 ↗Kurz gefasst
Fachartikel in der Zeitschrift Review of Scientific Instruments, online erschienen am 19.11.2026. Verfasst von Kelin Wu und 2 weiteren Autor:innen. Die Arbeit nennt 31 Literaturangaben. Gefördert von: National Natural Science Foundation of China.
Abstract (Originalsprache)
Defect detection on train wheel treads is a critical link in ensuring railway transportation safety. Existing deep learning detection models often suffer from scale misalignment and feature information loss during feature fusion, leading to insufficient perception of low-contrast edges in early-stage defects. To address these issues, this paper proposes a train wheel tread defect detection network based on a feature focusing diffusion pyramid network and dynamic task alignment. The network employs StarNet as its backbone to reduce computational cost for deployment on embedded devices. To resolve the problems of scale mismatch and insufficient semantic expression capability among feature maps from different levels, a texture and boundary aware focus module is designed to effectively enhance feature representation. A scale-adaptive dynamic feature extraction module, C3k2-SD-DWR, is adopted to improve the model’s capability to extract features of targets at different scales. The model integrates a task align dynamic detection head to enhance the spatial adaptability of features for tread defects. Experimental results on public datasets show that the proposed method achieves an average recognition accuracy of 95.41%, which is 4.9% higher than that of the original model, while reducing model size and computational cost by 1.8 MB and 0.3 GFLOPS, respectively, providing an efficient solution for automated detection of train wheel tread defects.
Förderung
- National Natural Science Foundation of China
Zitieren
APA
Wu, K., Wang, D., & Yao, Y. (2026). FDPT-Net (feature focusing diffusion pyramid and task-aligned network): A train wheelset tread detection network based on feature focusing diffusion pyramid and dynamic task alignment. Review of Scientific Instruments, 97(11). https://doi.org/10.1063/5.0336077
Vancouver
Wu K, Wang D, Yao Y. FDPT-Net (feature focusing diffusion pyramid and task-aligned network): A train wheelset tread detection network based on feature focusing diffusion pyramid and dynamic task alignment. Review of Scientific Instruments. 2026;97(11). doi:10.1063/5.0336077
BibTeX
@article{Wu2026,
author = {Wu, Kelin and Wang, Dan and Yao, Yupeng},
title = {FDPT-Net (feature focusing diffusion pyramid and task-aligned network): A train wheelset tread detection network based on feature focusing diffusion pyramid and dynamic task alignment},
journal = {Review of Scientific Instruments},
year = {2026},
volume = {97},
number = {11},
doi = {10.1063/5.0336077}
}
EUVIMED