Railway Track Crack Detection: A Comparative Study On Yolov7 And U-Net In Automated Inspection

Title

Railway Track Crack Detection: A Comparative Study On Yolov7 And U-Net In Automated Inspection

Creator

Vishal, S.; Shanthan, Hubert; Vijay Arputharaj, J.

Description

For railway networks to remain operationally safe and avoid catastrophic failures, structural integrity is essential. Track cracks can be found using labor-intensive, slow, and human error-prone manual inspection techniques. In this work, two cutting-edge deep learning models - YOLOv8 andU-Net v2 - for automated railway track crack detection using high-resolution imagery from Unmanned Aerial Vehicles (UAVs) are compared. In a real-world inspection scenario, we compare the different strategies of precise semantic segmentation (U-Net) and real-time object detection (YOLOv8) in order to assess their relative trade-offs. We compare performance on important metrics such as precision, recall, intersection over union (IoU), and inference speed using a custom dataset that was taken by a DJI Matrice 300 RTK drone. This work is novel because it examines how each model's output - bounding boxes versus pixel-level masks - directly affects the usefulness for maintenance workflows from an application-focused perspective. According to our research, U-Net v2 offers the fine-grained information required for precise damage assessment, while YOLOv8 is best suited for quick, extensive screening. This study offers railway operators useful information for creating a multi-stage, hybrid inspection strategy that strikes a balance between accuracy and speed. 2025 IEEE.

Source

Proceedings of the 2025 International Conference on Computational Innovations and Sustainable Technologies, ICCIST 2025;

Date

01-01-2025

Publisher

Institute of Electrical and Electronics Engineers Inc.

Subject

Computer Vision; Crack Detection; Deep Learning; DJI Matrice 300 RTK; Drone Inspection; Infrastructure Monitoring; Railway Safety; U-Net v2; UAV; YOLOv8

Coverage

Vishal S., Christ University, Department of Computer Science, Bengaluru, India; Shanthan H., Christ University, Department of Computer Science, Bengaluru, India; Vijay Arputharaj J., Christ University, Department of Computer Science, Bengaluru, India

Rights

Restricted Access; Hardcopy may be available in the library

Relation

ISBN: 979-833159676-7;

Format

online

Language

English

Type

Conference paper

Identifier

https://doi.org/10.1109/ICCIST67338.2025.11438571

https://www.scopus.com/pages/publications/105037441912?origin=resultslist