Algorithm for pixel-level concrete pavement crack segmentation based on an improved U-Net model.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Concrete surface cracks pose risks to road safety and infrastructure integrity.
• Accurate crack identification is essential for effective road condition assessment and maintenance planning.
• Existing crack detection methods often lack the required accuracy and speed for complex environments.

Purpose of the Study:

• To develop an improved U-shaped convolutional network (U-Net) model for enhanced concrete crack detection.
• To increase the accuracy, speed, and robustness of pavement crack identification algorithms.
• To broaden the predictive ability of crack detection models by training on diverse datasets.

Main Methods:

• Modification of the original U-Net model with two key innovations.
• Reduction of model parameters to improve efficiency.
• Training the enhanced U-Net-FML model on a combined dataset of public images and 960 custom road crack images.
• Utilizing datasets with varied exposure and noise conditions to enhance model generalization.

Main Results:

• The proposed U-Net-FML model demonstrated high accuracy and detection speed in complex environments.
• Achieved Mean Intersection over Union (MIoU) of 76.4%, F1 score of 74.2%, precision of 84.2%, and recall of 66.4%.
• Outperformed seven other comparison models in overall performance for crack detection tasks.

Conclusions:

• The U-Net-FML model offers significant engineering value for precise and efficient analysis of concrete pavement cracks.
• The enhanced model provides a robust solution for road condition assessment and maintenance strategy formulation.
• The study highlights the potential of deep learning for improving infrastructure monitoring and safety.