Automated Vision-Based Detection of Cracks on Concrete Surfaces Using a Deep Learning Technique

Area of Science:

• Computer Vision
• Structural Health Monitoring
• Artificial Intelligence

Background:

• Current computer vision crack detection methods struggle in on-site conditions.
• Visual inspection remains the standard despite limitations.

Purpose of the Study:

• To propose an automated crack morphology detection technique for concrete surfaces in on-site environments.
• To enhance structural inspection efficiency and accuracy.

Main Methods:

• Utilized a convolutional neural network (CNN), specifically AlexNet, trained on diverse image datasets.
• Implemented a five-class categorization including cracks, intact surfaces, similar patterns, and plants.
• Developed a probability map using a softmax layer for robust sliding window detection.

Main Results:

• Achieved high accuracy in crack detection on field images and real-time video frames from unmanned aerial vehicles.
• Demonstrated successful categorization of surface conditions, including subtle crack patterns.
• Validated the robustness and applicability of the method in challenging on-site environments.

Conclusions:

• The proposed CNN-based automated detection technique is highly adoptable for on-site crack inspection.
• This method offers a reliable alternative to traditional visual inspection for structural management.
• The technique shows significant potential for improving the safety and maintenance of concrete structures.