A Binocular Vision-Based Crack Detection and Measurement Method Incorporating Semantic Segmentation.

Area of Science:

• Civil Engineering
• Computer Vision
• Materials Science

Background:

• Concrete crack morphology is vital for assessing component condition.
• Existing deep learning methods for crack detection face limitations with monocular devices and fixed viewpoints.
• Robotic or UAV-based inspections are hindered by restrictions in image acquisition.

Purpose of the Study:

• To develop a non-contact, flexible, efficient, and accurate method for detecting and measuring concrete cracks.
• To overcome the limitations of monocular vision in crack dimension quantification.
• To enable 3D reconstruction of crack morphology for comprehensive analysis.

Main Methods:

• A binocular stereo vision system combined with a full convolutional network (FCN) for crack detection and segmentation.
• Utilizing an encoder-decoder architecture within the FCN for precise edge detail and segmentation.
• Implementing a central projection scheme for reliable 3D reconstruction of crack morphology, mitigating view mismatches.

Main Results:

• The FCN achieved high performance in crack segmentation on complex backgrounds, with a precision of 83.85%, recall of 85.74%, and F1 score of 84.14%.
• Binocular stereo vision enhanced image acquisition flexibility and streamlined the process.
• The 3D reconstruction demonstrated practical feasibility, with relative crack width measurement errors ranging from -3.9% to 36.0%.

Conclusions:

• The proposed binocular stereo vision and FCN method provides a robust solution for non-contact concrete crack inspection.
• This approach enhances the accuracy and flexibility of crack measurement, particularly for 3D morphology.
• The method shows practical feasibility for real-world infrastructure health monitoring.