FF-DEIM: DEIM with Image Dehazing and Self-Supervised Pretraining for Catenary Support Component Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Railway Engineering

Background:

• Catenary support components (CSCs) are critical for electrified railway safety.
• Existing image-based inspection methods struggle with multi-scale/multi-class objects and degraded image quality due to environmental factors (fog, rain, low light).
• These challenges hinder accurate detection and impact railway operational safety.

Purpose of the Study:

• To propose a novel detection framework, FF-DEIM, for robust catenary support component detection.
• To address challenges of multi-scale objects and poor image quality in railway inspection.
• To improve the accuracy and reliability of automated catenary inspection systems.

Main Methods:

• Introduced a dual-channel fusion network (DCFNet) to enhance image quality by removing foreground interferences.
• Developed a contrastive learning pretraining framework (MIMCL) for improved feature extraction and faster convergence.
• Proposed a feature-focusing pyramid network (FFPN) to enhance small object detection by fusing cross-level contextual features.

Main Results:

• The proposed DCFNet effectively improves image quality, mitigating issues from fog, raindrops, and blur.
• MIMCL pretraining enhances model focus on critical catenary components, optimizing feature extraction.
• FFPN demonstrates improved performance in detecting small catenary support components, validated on a diverse drone-based dataset.

Conclusions:

• The FF-DEIM framework offers a significant advancement in automated catenary support component detection.
• The method demonstrates robustness in challenging environmental conditions, including fog, rain, and low light.
• This research contributes to enhanced railway operational safety through improved inspection technologies.