Enhancing the reliability of building crack detection using convolutional neural networks via leveraging robust dataset design

The detection of cracks is important in the maintenance of structures like concrete and brick walls, because the appearance of cracks is considered an initial sign of deterioration of structures, ensuring the safety and durability of structures. Traditionally, crack detection is performed by a maintenance engineer manually, which is laborious and time-consuming. Structural maintenance has seen the emergence of automated crack detection methods as a major goal. Convolutional neural network (CNN)-based methods have been superior to other existing methods. But they are not always good in different environments, like shadows, colour changes, or noise, and only work well if the training data is labelled correctly. Thus, CNN-based crack detection requires high-quality labelled datasets. In this research, we assembled comprehensive datasets (captured and online) and employed them in CNN-based techniques (e.g., AlexNet, ResNet-50, GoogLeNet, and VGG16), followed by a comparative analysis to evaluate their performance in structural maintenance. In comparing the performance of the AlexNet, ResNet-50, GoogLeNet, and VGG16 models for crack detection in buildings, ResNet-50 emerged as the top-performing model. All four models achieved high accuracy; however, ResNet-50 consistently demonstrated superior precision, recall, and F1-score. With a test accuracy of 99.88% for ResNet-50, 99.56% for GoogLeNet, 99.25% for VGG16, and 95.31% for AlexNet, ResNet-50 proved more adept at interpreting complex data patterns and minimizing classification errors. This highlights ResNet-50’s stronger ability to enhance classification performance, positioning it as a preferred model for structural crack identification.