Post-disaster damage assessment plays a critical role, particularly following natural events such as earthquakes, where buildings and structures may incur damages of varying severity. Accurate classification of the damage level, ranging from minor to moderate to severe, is essential for assessing structural integrity, in order to recalibrate structural element strength coefficients, and guiding repair and reconstruction decisions. In recent years, Convolutional Neural Networks (CNNs) have emerged as powerful tools for image analysis and classification, opening new possibilities in post-earthquake damage assessment. The application of CNNs for damage assessment in civil engineering represents a relatively novel yet promising research area, by enabling accurate and rapid classification of structural damage. In this work, we focus on the comparison of several CNN architectures, such as GoogLeNet, VGG19, ResNet50, MobileNet, and others, by examining the design characteristics of these networks, as well as the results obtained through benchmark experiments. Experiments include an analysis of model performance under different conditions, such as the use of data augmentation and transfer learning, different learning rates, and different metrics, such as accuracy, recall, precision, and F1-score, training the models on the PEER Hub ImageNet ( \(\Phi \) -Net) dataset. Our goal is to identify which CNN architecture is the most suitable for this specific application, thereby contributing to the improvement of structural damage assessment methodologies.