Bridging Convolutional Neural Networks and Transformers for Efficient Crack Detection in Concrete Building...

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Accurate crack detection in buildings is crucial for safety, longevity, and economic value.
• Conventional Convolutional Neural Network (CNN) methods for crack detection have limitations, including high computational costs and shallow feature extraction.
• Existing deep learning (DL) techniques may not fully capture complex crack characteristics.

Purpose of the Study:

• To introduce a novel Convolution and Composite Attention Transformer Network (CCTNet) model for enhanced crack detection in building structures.
• To address the limitations of conventional CNNs, such as high computational costs and inadequate feature representation.
• To improve the accuracy, efficiency, and reliability of crack identification in the built environment.

Main Methods:

• Developed a novel CCTNet model integrating convolution, channel attention, and window-based self-attention mechanisms.
• Utilized an improved cross-attention module to enhance feature interaction and integration across windows.
• Leveraged both localized feature extraction (CNN) and global contextual understanding (self-attention).

Main Results:

• CCTNet achieved high precision rates: 98.60% on Historical Building Crack2019, 98.93% on SDTNET2018, and 99.33% on the proposed DS3 dataset.
• The model demonstrated near-zero training validation loss, indicating effective learning.
• Achieved Area Under the Curve (AUC) scores of 0.99 for Historical Building Crack2019 and 0.98 for SDTNET2018.

Conclusions:

• The proposed CCTNet model significantly outperforms existing methodologies for building crack detection.
• CCTNet offers an accurate, efficient, and reliable solution for identifying structural cracks.
• The model sets a new benchmark for automated crack detection in the built environment.