Modification and Evaluation of Attention-Based Deep Neural Network for Structural Crack Detection

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Structural health monitoring is crucial for safety, but traditional methods are inefficient and risky.
• Deep learning offers a promising solution for automated crack detection and recognition.

Purpose of the Study:

• To develop an improved deep learning model for accurate and efficient structural crack identification.
• To enhance the U-Net network using an attention mechanism for better crack recognition performance.

Main Methods:

• Comparative analysis of U-Net and lrassp models for crack detection.
• Improvement of the U-Net network by incorporating an attention mechanism, resulting in the ECA-UNet model.
• Experimental evaluation of the ECA-UNet model's performance on large-scale structural crack recognition tasks.

Main Results:

• The U-Net model demonstrated superior performance over the lrassp model.
• The proposed ECA-UNet model achieved a 0.016 increase in Intersection over Union (IOU) and a 0.131 increase in recall compared to the original U-Net.
• ECA-UNet exhibited superior crack recognition performance, accurately identifying cracks with minimal noise misidentification.

Conclusions:

• The ECA-UNet model significantly enhances structural crack identification accuracy and efficiency.
• Attention-based U-Net improvements are effective for real-world structural health monitoring applications.
• The proposed model offers a robust and reliable solution for large-scale crack recognition in structures.