Crack Detection in Images of Masonry Using CNNs.

Area of Science:

• Computer Vision
• Machine Learning
• Structural Health Monitoring

Background:

• Extensive research exists on crack detection in concrete and asphalt using computer vision.
• Crack detection in masonry structures, particularly brick-and-mortar, remains less explored.
• Reliable crack detection is crucial for assessing the structural integrity of masonry buildings.

Purpose of the Study:

• To train and evaluate a convolutional neural network (CNN) for crack detection in masonry images.
• To compare the CNN's performance against simpler classifiers using handcrafted features.
• To assess domain adaptation capabilities from laboratory to real-world masonry crack detection.

Main Methods:

• A convolutional neural network (CNN) was trained on laboratory images of brick walls.
• The CNN's crack detection performance was evaluated on both laboratory and real-world internet images.
• Comparison involved traditional classifiers utilizing handcrafted features for crack identification.

Main Results:

• The CNN demonstrated superior domain adaptation from laboratory to real-world masonry crack detection compared to simpler models.
• Simpler classifiers performed better in the reverse domain adaptation (real-world to laboratory).
• Machine learning methods, including CNNs, are capable of detecting cracks in masonry images.

Conclusions:

• This research validates the use of machine learning for masonry crack detection.
• Findings offer insights into improving model reliability for domain adaptation in masonry crack analysis.
• Further work is needed to enhance CNN performance across different imaging domains for masonry.