HE-YOLOv5s: Efficient Road Defect Detection Network.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Engineering

Background:

• Road defects pose a significant global safety risk, leading to increased traffic accidents.
• Existing road defect detection models face a trade-off between accuracy and speed, with poor generalization.
• There is an urgent need for efficient and accurate road defect detection systems.

Purpose of the Study:

• To develop an optimized deep learning model for accurate and fast road defect detection.
• To improve upon the YOLOv5s benchmark model for enhanced performance in road defect identification.
• To address the limitations of existing models regarding speed, accuracy, and generalization.

Main Methods:

• Model optimization through parameter reduction via pruning and module removal.
• Implementation of an improved Spatial Pyramid Pooling-Fast (SPPF) module for enhanced feature fusion.
• Integration of an attention module to focus on critical road defect features.
• Strategic replacement of activation functions and sampling methods.

Main Results:

• The proposed model demonstrates faster Frames Per Second (FPS) compared to the baseline YOLOv5s.
• Achieved a 2.08% improvement in Mean Average Precision (MAP) on the Global Road Damage Detection Challenge (GRDDC) dataset.
• Reduced the model size by 6.07 MB, indicating increased efficiency.
• The optimized model shows improved generalization capabilities.

Conclusions:

• The optimized YOLOv5s model effectively balances detection speed and accuracy for road defects.
• The enhancements, including SPPF and attention modules, significantly boost performance.
• This model presents a viable and efficient solution for real-world road defect monitoring and management.