Road Defect Identification and Location Method Based on an Improved ML-YOLO Algorithm.

Area of Science:

• Computer Vision and Machine Learning
• Civil Engineering and Infrastructure Monitoring

Background:

• Manual road defect inspection is inefficient and lacks precise localization.
• Existing automated methods may not fully capture complex defect features or adapt to scale variations.

Purpose of the Study:

• To develop and evaluate an enhanced ML-YOLO algorithm for accurate road defect identification and localization.
• To improve upon the baseline YOLOv8 object detection framework for road defect analysis.

Main Methods:

• Refinement of the YOLOv8 object detection framework, including convolutional layers and spatial pyramid pooling.
• Integration of Convolutional Block Attention (CBA) for enhanced feature capture.
• Incorporation of Selective Kernel Networks (SKN) for adaptive feature extraction and an optimized target localization algorithm.

Main Results:

• The enhanced ML-YOLO algorithm achieved a detection accuracy of 0.841, recall of 0.745, and average precision of 0.817.
• Significant improvements over the baseline YOLOv8 model, with increases in accuracy (+0.13), recall (+0.117), and average precision (+0.116).
• Demonstrated robust generalization capabilities on public datasets, outperforming YOLOv8n in average precision and recall.

Conclusions:

• The proposed ML-YOLO algorithm provides a highly accurate and precise method for road defect detection and localization.
• This approach enhances real-time road monitoring, contributing to reduced traffic accident risks and extended roadway lifespan.
• The algorithm shows strong generalization, making it suitable for diverse road infrastructure monitoring applications.