Improving the Concrete Crack Detection Process via a Hybrid Visual Transformer Algorithm.

Area of Science:

• Civil Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Biannual inspections of aging US concrete bridges are resource-intensive.
• Traditional crack detection methods are time-consuming and costly.
• There is a critical need for efficient bridge inspection technologies.

Purpose of the Study:

• To evaluate computer vision models for enhancing concrete bridge inspection efficiency.
• To compare the performance of Visual Transformer (ViT) and Convolutional Neural Network (CNN) models for crack detection.
• To assess the impact of image enhancement algorithms on detection accuracy.

Main Methods:

• Applied state-of-the-art Visual Transformer (ViT) and Convolutional Neural Network (CNN) models.
• Utilized over 20,000 high-quality images from the Concrete Images for Classification dataset.
• Integrated ViT with various image enhancement detector algorithms for benchmarking.

Main Results:

• A custom-built CNN model achieved over 99% accuracy in concrete crack detection.
• The CNN model demonstrated significantly lower training time compared to the ViT model.
• ViT combined with image enhancement detectors showed improved concrete crack detection accuracy.

Conclusions:

• Computer vision, particularly CNNs, offers an efficient solution for concrete bridge crack detection.
• These advancements enhance infrastructure safety, support resource conservation, and align with Industry 4.0 automation goals.
• Automating manual inspections reduces costs and integrates technology into public infrastructure management.