SPR-DETR: DETR with Self-Supervised Learning and Position Relation Modeling for UAV-Based Catenary Support Component...

Area of Science:

• Railway Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Catenary Support Components (CSCs) are critical for electrified railways, but their detection in images is challenging due to limited labeled data, complex backgrounds, and scale variations.
• Existing detection methods struggle with the scarcity of annotated data and the diverse visual conditions encountered in real-world railway environments.

Purpose of the Study:

• To develop a robust and efficient framework for detecting Catenary Support Components (CSCs) in railway images.
• To address the challenges of limited labeled data, complex backgrounds, and multi-scale variations in CSC detection.
• To reduce the cost and effort associated with data annotation for railway infrastructure monitoring.

Main Methods:

• A Siamese-based self-supervised learning framework was employed for pre-training to leverage unlabeled data and minimize annotation requirements.
• Novel modules, Vision Attention-based Intrascale Feature Interaction (Vision-AIFI) and Relation Vision Module (RVM), were introduced to enhance multi-scale feature extraction and handle background complexities.
• A Dempster-Shafer (DS) evidence theory-based detection head was integrated to improve classification confidence and localization precision.

Main Results:

• The proposed framework achieved high detection performance, with mAP of 77.84, APs of 67.84, APm of 70.31, and APl of 90.04 on a newly constructed UAV-based CSC dataset.
• The self-supervised pre-training significantly reduced the need for labeled data, lowering annotation costs.
• Domain generalization experiments demonstrated the framework's strong adaptability and high detection accuracy in real-world scenarios.

Conclusions:

• The developed framework effectively addresses the challenges in Catenary Support Component detection, offering a cost-effective and accurate solution.
• The integration of self-supervised learning, advanced feature interaction modules, and DS evidence theory provides a powerful tool for automated railway infrastructure inspection.
• The study highlights the potential of deep learning approaches for enhancing the safety and efficiency of railway operations through improved component monitoring.