LRD-DETR: A Lightweight RT-DETR-Based Model for Road Distress Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Pavement distress detection faces challenges with irregular shapes, overlooked fine cracks, and high model complexity.
• Existing methods struggle with accuracy and practical deployment in highway maintenance.

Purpose of the Study:

• To propose a lightweight road distress detection model (LRD-DETR) for improved accuracy and practicality.
• To address limitations of current pavement distress detection technologies.

Main Methods:

• Integrated C2f-LFEM module and ADown adaptive down-sampling into the backbone network.
• Embedded frequency-domain spatial attention and polarity-aware linear attention in feature layers.
• Developed a cross-scale spatial feature fusion module (CSF²M) for robust multi-level feature fusion.

Main Results:

• LRD-DETR improved F1-score by 7.1% and mAP@50 by 9.0% compared to baseline RT-DETR.
• Reduced computational complexity by 43.8% and parameter quantity by 38.0%.
• Demonstrated suitability for real-time detection on resource-limited embedded platforms.

Conclusions:

• The proposed LRD-DETR offers a significant advancement in lightweight and accurate road distress detection.
• The model's efficiency and performance make it ideal for practical highway maintenance applications.
• Enhanced feature representation and fusion strategies contribute to robust detection under complex conditions.