Improved YOLOv8n-based bridge crack detection algorithm under complex background conditions

Area of Science:

• Computer Vision
• Artificial Intelligence
• Structural Engineering

Background:

• Deep learning image processing is widely used for bridge crack detection.
• Challenges include missed detections and false positives due to lighting, stains, and dense cracks.

Purpose of the Study:

• To propose an improved YOLOv8n model for enhanced bridge crack detection.
• To address limitations of existing methods in detecting challenging crack scenarios.

Main Methods:

• Incorporated global attention mechanism in Backbone and Neck for improved feature extraction.
• Optimized feature fusion using Gam-Concat and replaced FPN-PAN upsample with DySample.
• Integrated MPDIoU loss in the Head to refine bounding box regression for dense cracks.

Main Results:

• Achieved a 3.02% increase in mAP@0.5 and 3.39% in mAP@0.5:0.95 compared to the original YOLOv8n model.
• Demonstrated significant improvements in precision (2.26%) and recall (0.81%).
• Outperformed other comparative models in detection accuracy, especially for dense and varied-scale cracks.

Conclusions:

• The improved YOLOv8n model effectively enhances bridge crack detection accuracy and robustness.
• The proposed enhancements offer practical value for bridge inspection and structural health monitoring.