An improved transformer-based concrete crack classification method.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Surface cracks in concrete structures are critical indicators of durability and serviceability.
• Current convolutional neural networks (CNNs) for crack identification are computationally expensive and inefficient.

Purpose of the Study:

• To propose an optimized Cross Swin transformer-skip (CSW-S) network for accurate and efficient concrete crack classification.
• To improve upon existing transformer and CNN models for crack detection.

Main Methods:

• Developed a novel CSW-S network by incorporating residual links into the Cross Swin transformer architecture.
• Trained and evaluated the CSW-S model on a dataset comprising 17,000 concrete crack images.

Main Results:

• The enhanced CSW-S network demonstrated an improved ability to extract image features, leading to higher crack recognition accuracy.
• Achieved a detection accuracy of 96.92% without pretraining.
• The CSW-S model exhibited superior recognition efficiency and accuracy compared to traditional transformer and CNN models.

Conclusions:

• The optimized CSW-S network offers a more accurate and efficient solution for concrete crack classification.
• This advancement contributes to better structural health monitoring and durability assessment in civil engineering applications.