Research on a Road Crack Detection Method Based on YOLO11-MBC.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Infrastructure Monitoring

Background:

• Existing pavement crack identification methods suffer from low accuracy and high false/missed detection rates, especially under complex road conditions.
• Pavement cracks present challenges due to their elongated topology and similarity to other road features like tree roots and lane markings.

Purpose of the Study:

• To develop a novel and accurate pavement crack identification method for complex road conditions.
• To enhance the detection of minute pavement cracks and reduce false positives and missed detections.

Main Methods:

• Proposed YOLO11-MBC model based on YOLO11, featuring a Multi-scale Feature Fusion Backbone Network (MFFBN) for improved feature extraction.
• Introduced BiMCNet, combining Bidirectional Feature Pyramid Network (BiFPN) with a Multimodal Cross-Attention (MCA) mechanism, to optimize detection of minute cracks.
• Implemented CGeoCIoU loss function to improve the alignment between predicted and ground-truth bounding boxes.

Main Results:

• YOLO11-MBC achieved a 22.5% improvement in F1-score and an 8% increase in mAP50 compared to the baseline YOLO11.
• The improved algorithm demonstrated superior performance against YOLOv8, YOLOv10, and YOLO11, with precision, recall, F1-score, mAP50, and mAP50-95 of 61%, 70%, 72%, 75%, and 66%, respectively.
• Validation conducted on the RDD2022 dataset confirmed the effectiveness of the proposed modules and overall approach.

Conclusions:

• The proposed YOLO11-MBC method significantly enhances pavement crack detection accuracy under complex road conditions.
• The integration of MFFBN, BiMCNet, and CGeoCIoU loss function effectively addresses limitations of existing methods.
• The study validates a robust and accurate approach for automated pavement crack identification.