Automatic Recognition of Road Damage Based on Lightweight Attentional Convolutional Neural Network

Area of Science:

• Computer Science
• Civil Engineering
• Artificial Intelligence

Background:

• Road defects pose risks to motorists and increase maintenance costs for authorities.
• Existing road damage detection systems may lack efficiency or accuracy for real-time applications.

Purpose of the Study:

• To propose a lightweight, end-to-end road damage detection network for fast, automatic, and accurate identification and classification of multiple road damage types.
• To improve detection accuracy and efficiency using a novel network architecture.

Main Methods:

• Developed a backbone network combining lightweight feature detection modules with a multi-scale feature fusion network.
• Integrated an embedded lightweight attention module to enhance feature information and improve detection accuracy with fewer parameters.
• Utilized vehicle-captured images for training and testing the road damage detection model.

Main Results:

• The proposed model demonstrates higher performance and fewer parameters compared to other representative models.
• The multi-scale feature fusion network aids in identifying road damage at various distances and angles.
• The embedded attention module enhances feature extraction, leading to improved detection accuracy.

Conclusions:

• The developed lightweight network efficiently identifies and classifies multiple road damage types from vehicle-mounted camera images.
• The system meets real-time detection requirements for mobile applications, offering practical benefits for road maintenance and safety.
• This approach provides a promising solution for automated road condition monitoring.