Deep Learning Classification of Concrete Crack Images using Pre-trained AlexNet and GoogleNet Models

Cracking in concrete walls is a common issue in construction, often leading to structural weakening. Such damage can significantly impact both maintenance and repair costs. Concrete wall damage can be categorized in various ways, and detecting these cracks is crucial for timely intervention. This study employed deep learning neural networks to accelerate and simplify the detection and classification of concrete crack images. Two deep learning models, AlexNet and GoogleNet, were trained to classify cracks according to three categories: severe cracks (class 1), fine cracks (class 2), and normal concrete (class 3), using a total of 10,601 concrete images. Among the models, GoogleNet performed than AlexNet. The average performance for GoogleNet is 94.04% for accuracy, 91.96% for precision, 91.97% for recall, 95.39% for specificity, and 91.77% for F-score. However, despite having a lower accuracy than GoogleNet, AlexNet exhibited faster training times, which may be advantageous in certain applications where processing speed is prioritized over precision. These results demonstrate that while GoogleNet provides higher accuracy, AlexNet offers a more time-efficient solution.