Ensemble model for rail surface defects detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Rail surface defect detection is crucial for high-speed rail safety and maintenance.
• Convolutional Neural Network (CNN) models are effective but require extensive training data.
• Existing CNN models vary in performance and are sensitive to data availability.

Purpose of the Study:

• To develop an ensemble framework for industrialized rail defect detection.
• To address the challenge of limited annotated data for training CNN models.
• To improve the performance and robustness of rail defect detection systems.

Main Methods:

• An ensemble framework combining multiple backbone networks for feature extraction.
• Binary feature mixing to create diverse sub-networks.
• Random application of image and feature augmentation techniques.
• Utilizing a shared feature pyramid network to optimize parameters and reduce computational cost.

Main Results:

• The proposed ensemble framework outperforms single CNN architectures for rail defect detection.
• Achieved a 7.4% higher mean Average Precision (mAP.5) compared to YOLOv5.
• Achieved a 2.8% higher mAP.5 compared to Faster R-CNN on a dataset with 8 defect classes.

Conclusions:

• The ensemble approach offers superior performance in rail defect detection tasks, especially with limited data.
• The framework enhances model diversity and robustness through feature mixing and augmentation.
• This method provides a more efficient and accurate solution for industrial rail maintenance.