Dual-Branch Cross-Fusion Normalizing Flow for RGB-D Track Anomaly Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Inspection

Background:

• 2D anomaly detection in railway inspection is limited by image acquisition conditions.
• Integrating depth maps with RGB data is explored to mitigate these interferences.
• Novel approaches are needed for robust multi-modal anomaly detection in industrial settings.

Purpose of the Study:

• To propose a novel RGB-D anomaly detection method for railway track inspection.
• To leverage dual-branch normalizing flow with multi-modal inputs for enhanced detection.
• To improve the accuracy and robustness of anomaly detection in challenging railway environments.

Main Methods:

• Developed Dual-Branch Cross-Fusion Normalizing Flow (DCNF) for RGB-D anomaly detection.
• Introduced a mutual perception module for early-stage cross-complementary knowledge acquisition.
• Implemented a fusion flow strategy to effectively integrate dual-branch RGB-D inputs.

Main Results:

• Achieved an impressive Area Under the Receiver Operating Characteristic Curve (AUROC) score of 98.49% on the Track Anomaly (TA) dataset.
• Demonstrated a performance improvement of 3.74% over the second-best method.
• Validated the effectiveness of the proposed fusion strategy for multi-modal anomaly detection.

Conclusions:

• DCNF offers a significant advancement in RGB-D anomaly detection for railway inspection.
• The proposed mutual perception and fusion flow modules enhance the detection of track anomalies.
• The method shows strong potential for real-world industrial inspection applications.