RailTrack-DaViT: A Vision Transformer-Based Approach for Automated Railway Track Defect Detection

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Railway track defects present significant safety risks and economic consequences.
• Manual inspection methods are labor-intensive, costly, and susceptible to human error.

Purpose of the Study:

• To develop and evaluate a novel vision transformer-based approach for automated railway track defect classification.
• To improve the accuracy and efficiency of defect detection compared to existing methods.

Main Methods:

• Proposed RailTrack-DaViT, a model utilizing the Dual Attention Vision Transformer (DaViT) architecture.
• Trained and evaluated the model on diverse datasets: rail, fastener and fishplate, multi-faults, and ThaiRailTrack.
• Compared performance against state-of-the-art Convolutional Neural Network (CNN)-based methods.

Main Results:

• Achieved high accuracies: 96.9% (rail), 98.9% (fastener/fishplate), 98.8% (multi-faults), and 99.2% (ThaiRailTrack).
• Demonstrated superior performance over CNN-based methods.
• Showcased quick adaptation to unseen data and improved model stability during fine-tuning.

Conclusions:

• RailTrack-DaViT offers a highly accurate and efficient solution for railway track defect classification.
• The model's adaptability and stability can significantly reduce inspection time in practical scenarios.
• Enhances railway safety through advanced AI-driven defect detection.