Pavement-DETR: A High-Precision Real-Time Detection Transformer for Pavement Defect Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Accurate road defect detection is vital for road maintenance safety and efficiency.
• Challenges include distinguishing foreground/background and identifying small or overlapping defects.

Purpose of the Study:

• To propose an optimized Pavement-DETR model for accurate road defect detection.
• To enhance detection of common pavement defects like cracks and potholes.

Main Methods:

• Introduced Channel-Spatial Shuffle (CSS) attention in ResNet backbone for precise feature focus.
• Employed Conv3XC structure for improved foreground-background differentiation and feature fusion.
• Developed a combined loss function using PIoU v2 and Normalized Wasserstein Distance (NWD) for comprehensive bounding box optimization.

Main Results:

• Pavement-DETR achieved a 7.7% mAP improvement at IoU=0.5 on the UAV-PDD2023 dataset.
• mAP increased by 8.9% at IoU=0.5-0.95, and F1 Score improved by 7% compared to baseline.
• Demonstrated superior performance in detecting small-sized and overlapping road defects.

Conclusions:

• The Pavement-DETR model significantly enhances road defect detection accuracy.
• The proposed improvements effectively address challenges in identifying diverse and complex pavement defects.
• This model offers substantial benefits for efficient and safe road maintenance operations.