|
|
Civil-Comp Conferences
ISSN 2753-3239
CCC: 15
PROCEEDINGS OF THE SEVENTH INTERNATIONAL CONFERENCE ON
RAILWAY TECHNOLOGY:
RESEARCH, DEVELOPMENT AND MAINTENANCE
Edited by: J. Pombo
Paper 17.5 A Novel Approach to Railway Wheel Defect Detection Using Custom Deep Learning Models for Monocular Depth Estimation M.Z. Shaikh 1,2,3, B. Abro 2,3, E.N. Baro 4, S. Jatoi 2,3, E.B. Blazquez-Parra 5, B.S. Chowdhry 2,3 and A. Kolmykova 61Mechanical Engineering and Energy Efficiency, School of Industrial Engineering, University of Malaga, Spain
2National Center for Robotics, Automation and Artificial Intelligence, Mehran University of Engineering and Technology (MUET), Jamshoro, Pakistan
3NCRA-CMS Lab, Mehran University of Engineering and Technology (MUET), Jamshoro, Pakistan
4Departamento de Ingenier´ıa de Comunicaciones, Campus de Teatinos, Universidad de Malaga, Spain
5Department of Graphical Engineering, Design and Projects, Universidad de Malaga, Spain
6, Zhytomyr Polytechnic State University, Ukraine Full Bibliographic Reference for this paper
M.Z. Shaikh, B. Abro, E.N. Baro, S. Jatoi, E.B. Blazquez-Parra, B.S. Chowdhry, A. Kolmykova, "A Novel Approach to Railway Wheel Defect Detection Using Custom Deep Learning Models for Monocular Depth Estimation", in J. Pombo, (Editor), "Proceedings of the Seventh International Conference on
Railway Technology:
Research, Development and Maintenance
",
Civil-Comp Press, Edinburgh, UK,
Online volume: CCC 15, Paper 17.5, 2026, doi:10.4203/ccc.15.17.5
Keywords: railway, depth estimation, computer vision, convolutional neural networks, U-Net, ResNet50, transformer.
Abstract
The monocular depth estimation in the field of computer vision represents significant challenges that are crucial in many applications such as autonomous driving and augmented reality. Recent development in deep learning addresses the challenges of depth prediction requiring both fine-grained details and structural integrity. This research introduces novel techniques that integrate the state-of-the-art convolutional neural networks with other advanced architectures including custom U-Net, ResNet50 and Vision Transformer to produce depth maps based on images. The evaluation of experiments showed that custom UNet outperformed other models with high accuracy and generalization than ResNet50 and ViT. It achieved significantly small RMSE of 0.1267 and loss of 0.0160 compared to other models that overfit the results with higher RMSE and loss scores.
download the full-text of this paper (PDF, 12 pages, 489 Kb)
go to the previous paper
go to the next paper
return to the table of contents
return to the volume description
|
