|
Title |
Effect of Data Size and Hyperparameter Settings on Classification Performance and Computational Efficiency in CNN-Based Concrete Damage Classification
|
|
Authors |
김일순(Il Sun Kim) ; 양은익(Eun Ik Yang) |
|
DOI |
https://doi.org/10.11112/jksmi.2026.30.2.52 |
|
Keywords |
계산 효율; 합성곱 신경망; 데이터 수; 하이퍼파라미터 최적화; 구조물 손상 분류 Computational efficiency; Convolutional neural network; Dataset size; Hyperparameter optimization; Structural damage classification |
|
Abstract |
This study comprehensively analyzed the effects of hyperparameter settings, dataset size, damage-type-specific performance characteristics, and computational efficiency on CNN-based structural damage classification. Experiments were conducted using four CNN models?GoogLeNet, ResNet-50, EfficientNet-B0, and MobileNetV2?and classification performance was quantitatively compared based on test-set accuracy and Macro F1-score. The results demonstrated that, even within the same model, classification accuracy and performance variability varied substantially depending on the hyperparameter configurations. Increasing the training dataset size generally improved overall performance while mitigating sensitivity to hyperparameter variations. However, beyond a certain dataset size, performance gains tended to saturate, and the extent of this saturation differed depending on the model architecture. In addition, under balanced class distributions, performance differences among damage types were minimized. A clear trade-off between classification performance and computational efficiency was observed across the models. This study contributes to the literature by providing an integrated analysis of hyperparameter sensitivity, data efficiency, and computational cost, rather than focusing solely on single optimal performance comparisons.
|