CNN-Based Crack Detection of Reinforced Concrete Slab Culverts

Concrete structures are a vital component of modern infrastructure, but they are subjected to cracking over time due to environmental and structural factors. In addition to allowing access to dangerous and corrosive chemicals in concrete, cracks also allow water to seep through the structural members, accelerating the corrosion of reinforcement and damaging the aesthetics of the structure. Traditional visual inspection methods are time-consuming and expensive and often result in subjective interpretations, thereby necessitating automated crack detection techniques. Convolutional Neural Networks (CNNs) are a type of Deep Learning algorithm that has shown remarkable performance in image classification, object detection, and segmentation tasks. Transfer Learning can fine-tune pre-trained CNN models like GoogLeNet, VGGNet and ResNet with high accuracy and reduced overfitting, even with a limited training dataset. In this study, ResNet is fine-tuned using publicly available datasets of images of cracks on concrete surfaces. Subsequently, the trained model is used to detect cracks on a 1:8 scale reinforced concrete slab culvert built in the laboratory and subjected to bidirectional seismic excitation on a shake table. OpenCV has been adopted to extract features, such as crack width, orientation, and shape, from the detected cracks, which can be used for further analysis, such as crack classification, severity assessment and damage evaluation. The key findings from this study indicate that CNN coupled with OpenCV can be a viable alternative to manual crack inspection and analysis tasks. The application can be extended to help structural engineers identify the condition and extent of damage in large-scale structures, like bridges and dams, using camera-equipped unmanned aerial systems (UAS).