https://doi.org/10.11112/jksmi.2026.30.4.166
김일순(Il Sun Kim) ; 양은익(Eun Ik Yang)
This study compared the classification performance and repeated training stability of convolutional neural network (CNN) models for concrete damage image classification under different optimizer and initial learning rate conditions. The target damage types consisted of three classes: crack, efflorescence, and rebar exposure. Four transfer learning-based CNN models, namely EfficientNet-B0, GoogLeNet, MobileNetV2, and ResNet-50, were applied. Adam, RMSProp, and SGDM were used as optimizers, and the initial learning rates were set to 0.001, 0.0003, and 0.0001. For each model?optimizer?learning rate combination, repeated training was performed by changing the random seed 10 times, resulting in a total of 360 training runs. The performance was evaluated using the Test Macro F1-score, standard deviation, minimum value, coefficient of variation, and computational time. The results showed that all CNN models achieved an average Test Macro F1-score higher than 0.94, while the differences in average performance among the models and optimizers were limited. However, performance degradation and repeated training variability were observed in specific model?optimizer?learning rate combinations. In terms of damage type, the rebar exposure class showed the highest average F1-score, whereas the efflorescence class showed relatively lower performance. Therefore, in CNN-based concrete damage classification, models and training conditions should be selected by considering not only average performance but also repeated training stability, minimum performance, class-wise performance, and computational efficiency.