콘크리트 구조물 균열 탐지 자동화를 위한 VGG-T 이미지 분류 모델 개발

초록

Crack detection plays a crucial role in monitoring and inspecting the condition of construction structures. Traditional Convolutional Neural Network (CNN) methods, which focus mainly on local feature extraction, face limitations in accuracy. In contrast, Vision Transformer (ViT) models effectively capture global features but require large-scale datasets for training. To overcome these challenges, the VGG-T Image Classification model is proposed. This model combines the local feature extraction strength of the CNN-based VGG-16 with the global feature learning capabilities of ViT. Incorporating transfer learning and data augmentation techniques allows effective training even with small datasets. The model was evaluated using binary classification metrics and compared against VGG-16, VGG-19, ResNet-101, and ViT models. Results showed an accuracy of 99.6%, demonstrating that integrating these two architectures significantly improves detection accuracy. This advancement is expected to contribute to the development of structural safety diagnosis, automated safety maintenance, and crack detection technologies.

키워드