SDC-YOLOv8: An Improved Algorithm for Road Defect Detection Through Attention-Enhanced Feature Learning and Adaptive...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Infrastructure Management

Background:

• Existing object detection algorithms struggle with small road defects and complex environmental conditions.
• Inaccurate road defect detection compromises traffic safety and timely repairs.

Purpose of the Study:

• To develop an improved YOLOv8-based algorithm for accurate and efficient road defect detection.
• To enhance the detection of small road surface defects under challenging lighting and background scenarios.

Main Methods:

• Proposed SDC-YOLOv8 algorithm integrating SPPF-LSKA module for multi-scale feature representation and DySample dynamic upsampling for adaptive feature reconstruction.
• Incorporated Coordinate Attention module to boost spatial localization accuracy in complex conditions.
• Utilized Fast Spatial Pyramid Pooling with Large Separable Kernel Attention (SPPF-LSKA) and DySample dynamic upsampling.

Main Results:

• SDC-YOLOv8 achieved 78.0% mAP@0.5, 81.0% Precision, and 70.7% Recall on a public pothole dataset.
• The model demonstrated real-time performance at 85 FPS, outperforming the baseline YOLOv8n model.
• Achieved a 2.0 percentage point improvement in mAP@0.5, 3.3% in Precision, and 1.8% in Recall, with an F1 score of 75.5%.

Conclusions:

• SDC-YOLOv8 effectively enhances small-target detection accuracy for road defects.
• The algorithm maintains real-time processing capabilities, offering a practical solution for intelligent road defect detection.
• The proposed method provides a significant advancement for road maintenance and traffic safety applications.