A Lightweight Pavement Defect Detection Algorithm Integrating Perception Enhancement and Feature Optimization.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Civil Engineering

Background:

• Pavement defect detection models face challenges with large computations and balancing complexity with accuracy.
• Existing methods struggle to efficiently extract and fuse multi-scale features for precise defect identification.

Purpose of the Study:

• To develop a lightweight pavement defect detection algorithm, PGS-YOLO, that overcomes computational limitations and improves detection accuracy.
• To enhance feature extraction and fusion capabilities for more effective pavement defect identification.

Main Methods:

• Proposed PGS-YOLO algorithm based on YOLOv8, integrating perception enhancement and feature optimization.
• Designed RFCBAMConv and C2f-RFCBAM for the Perception Enhanced Feature Extraction Network (PEFNet) to improve multi-scale feature extraction.
• Introduced DySample for the Generalized Dynamic Sampling Feature Pyramid Network (GDSFPN) to optimize feature fusion.
• Utilized a shared detail-enhanced convolution lightweight detection head (SDCLD) and Wise-IoU for improved performance.

Main Results:

• PGS-YOLO achieved a 2.8% and 2.9% increase in mAP50 on the GRDDC2022 dataset and its Chinese subset, respectively.
• Reduced parameters by 10.3% and computation by 9.9% compared to YOLOv8n, with a 69 FPS frame rate.
• Improved mAP50 by 2.3% on the CRACK500 dataset, demonstrating a better balance between complexity and accuracy.

Conclusions:

• PGS-YOLO offers a superior balance between model complexity and detection accuracy for pavement defect detection.
• The algorithm provides efficient, real-time performance suitable for practical pavement inspection applications.
• The proposed enhancements in feature extraction and fusion contribute to state-of-the-art results in pavement defect detection.