Crack Unet: Crack Recognition Algorithm Based on Three-Dimensional Ground Penetrating Radar Images

Area of Science:

• Geophysics
• Civil Engineering
• Artificial Intelligence

Background:

• Three-dimensional ground-penetrating radar (3D GPR) is effective for detecting internal pavement cracks.
• Manual interpretation of 3D GPR images is inefficient and requires significant personnel.
• Current limitations hinder the widespread adoption of 3D GPR for pavement analysis.

Purpose of the Study:

• To propose an improved Crack Unet model for automated crack segmentation in 3D GPR images.
• To enhance the efficiency and accuracy of pavement internal crack detection.
• To overcome the limitations of manual interpretation and high personnel requirements in 3D GPR data analysis.

Main Methods:

• Development of an improved Crack Unet model based on the Unet semantic segmentation architecture.
• Integration of novel modules (PMDA, MC-FS, RS) to enhance segmentation performance.
• Comparative analysis against mainstream algorithms like deepLabv3, PSPNet, and Unet.

Main Results:

• The Crack Unet model demonstrated improved Mean Pixel Accuracy (MPA) and Mean Intersection over Union (MioU).
• It outperformed deepLabv3, PSPNet, and Unet in radar image crack segmentation tasks.
• Ablation studies confirmed the effectiveness of integrated modules, particularly PMDA, over SE and CBAM modules.

Conclusions:

• The Crack Unet model significantly improves crack segmentation performance in 3D GPR images.
• It offers a more efficient and accurate alternative to manual interpretation for pavement damage assessment.
• The model exhibits excellent engineering application value for pavement crack detection.