Shield Tunnel Crack Detection Based on Improved Unet

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Unet is a popular deep learning model for crack detection in tunneling.
• Existing Unet models face limitations in feature expression and segmentation accuracy.

Purpose of the Study:

• To propose DTA-Unet, an improved deep learning model for enhanced crack detection.
• To address limitations in feature extraction and segmentation accuracy of conventional Unet models.

Main Methods:

• DTA-Unet utilizes dynamic convolution decomposition (DCD) to boost feature extraction.
• Triple attention (TA) combined with attention gate (AG) refines segmentation by reducing redundant spatial and channel information.
• The model was evaluated on crack datasets against Unet, image processing algorithms, and other deep neural networks.

Main Results:

• DTA-Unet demonstrated superior performance in crack detection compared to conventional Unet and other advanced methods.
• The enhanced feature extraction and attention mechanisms led to more accurate crack segmentation.

Conclusions:

• DTA-Unet offers significant improvements for shield tunnel crack detection and maintenance.
• The proposed DCD and TA mechanisms effectively enhance deep learning-based crack segmentation.