Comparative study of CNN-based deep learning models for concrete crack detection

초록

This study aims to systematically compare the performance of nine convolutional neural network (CNN) models-AlexNet, GoogLeNet, VGG16, VGG19, ResNet-18, ResNet-50, ResNet-101, SqueezeNet, and MobileNetV2-for automated concrete crack detection, addressing the growing need for reliable and efficient inspection in infrastructure maintenance. A custom dataset comprising 3000 training and 12,000 validation images (256 x 256 pixels) was constructed, along with five high-resolution test images (4160 x 3120 pixels) captured under diverse environmental conditions. Transfer learning was applied to fine-tune pre-trained models using MATLAB, and the models were evaluated based on six performance metrics. A sliding window technique was employed to generate crack maps for visual assessment. The results showed that VGG16 and VGG19 consistently outperformed other models across all metrics and demonstrated robust generalization to lighting and surface variations. The study contributes practical insights into model selection for real-world applications and highlights the feasibility of using lightweight training data. This deployment-oriented, comparative evaluation under realistic constraints represents a novel contribution to the field of automated crack detection.

키워드