Research on steel rail surface defects detection based on improved YOLOv4 network

Area of Science:

• Computer Vision
• Artificial Intelligence
• Materials Science

Background:

• Steel rail surface image acquisition is challenging due to light variations and background clutter.
• Detecting and recognizing rail defects is crucial for railway safety and maintenance.

Purpose of the Study:

• To develop a deep learning algorithm for accurate and efficient detection of rail defects.
• To address challenges like inconspicuous defect edges, small defect sizes, and background interference.

Main Methods:

• Sequential application of rail region extraction, improved Retinex image enhancement, background modeling difference, and threshold segmentation.
• Integration of Res2Net and CBAM attention mechanisms for improved feature extraction and small target detection.
• Modification of the PANet structure by removing the bottom-up path enhancement to reduce redundancy.

Main Results:

• Achieved an average accuracy of 92.68% and a recall rate of 92.33% for rail defect detection.
• Demonstrated an average detection time of 0.068 seconds per image, meeting real-time detection requirements.
• The improved YOLOv4 model outperformed mainstream algorithms like Faster RCNN, SSD, and YOLOv3 in key performance metrics (P, R, F1).

Conclusions:

• The proposed deep learning method significantly improves rail defect detection accuracy and efficiency.
• The enhanced YOLOv4 model shows excellent comprehensive performance and is suitable for practical rail defect detection projects.