Enhanced Rail Surface Defect Segmentation Using Polarization Imaging and Dual-Stream Feature Fusion.

Area of Science:

• Computer Vision
• Materials Science
• Industrial Safety

Background:

• Traditional visual inspection struggles with low-contrast rail surface defects in poor lighting.
• Existing methods fail to effectively detect small defects against complex backgrounds.

Purpose of the Study:

• To develop a novel defect segmentation method for rail surfaces.
• To enhance the detection of low-contrast and small defects, especially in challenging lighting conditions.

Main Methods:

• A dual-stream feature fusion network combining polarization images with DeepLabV3+.
• Utilized pruned MobileNetV3 backbone with coordinate attention for efficiency.
• Integrated CBAM in the decoding stage for refined feature fusion.

Main Results:

• Achieved 73.00% MIoU and 80.59% MPA, outperforming U-Net, PSPNet, and original DeepLabV3+.
• Demonstrated a comprehensive detection accuracy of 97.82%.
• Effectively detected small, low-contrast defects in complex backgrounds.

Conclusions:

• The proposed dual-stream network with polarization imaging significantly improves rail surface defect detection.
• The method meets stringent accuracy requirements for industrial safety and efficiency.
• This approach offers a robust solution for challenging defect detection scenarios.