Optimized YOLOv11m for real-time high-speed railway catenary defect detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Real-time defect detection in high-speed railway catenary systems is crucial for safety and maintenance.
• Existing object detection models, like YOLO, face challenges with small components and high computational demands.
• Limitations include difficulty in detecting small parts (e.g., cotter pins) and platform constraints.

Purpose of the Study:

• To develop an optimized object detection model for real-time inspection of railway catenary components.
• To address the limitations of existing models in detecting small objects and managing computational costs.
• To propose a lightweight and accurate solution for automated defect detection.

Main Methods:

• An optimized YOLOv11m model, termed MSIM-YOLOv11m, was developed.
• Integration of three novel modules: Large Separable Kernel Attention (LSKA), Bidirectional Feature Pyramid Network (BiFPN), and Adaptive Kernel Convolution (AKConv).
• Model evaluated on a dedicated catenary dataset for defect detection performance.

Main Results:

• The MSIM-YOLOv11m model achieved a mAP50-95 of 78.3% and a small-target AP of 64.7%.
• Demonstrated a 50.5% reduction in computational cost compared to the YOLOv9m model.
• The model proved effective in detecting small-sized defects on catenary components.

Conclusions:

• MSIM-YOLOv11m offers a lightweight and accurate solution for real-time railway catenary inspection.
• The proposed model effectively overcomes challenges associated with small object detection and computational efficiency.
• This advancement supports enhanced safety and maintenance through automated visual inspection.