|
Title |
Repeated-Training-Based Evaluation of Classification Performance and Training Stability of CNN Models for Concrete Damage Classification
|
|
Authors |
김일순(Il Sun Kim) ; 양은익(Eun Ik Yang) |
|
DOI |
https://doi.org/10.11112/jksmi.2026.30.4.17 |
|
Keywords |
계산 효율성; 합성곱신경망(CNN); 이미지 분류; 학습 안정성; 반복 학습 평가 Computational efficiency; Convolutional Neural Network; Image classification; Training stability; Repeated training evaluation |
|
Abstract |
This study evaluated the classification performance, training stability, and computational efficiency of four CNN models?EfficientNet, MobileNet, ResNet, and GoogLeNet?for concrete structural damage image classification using a repeated-training framework. To overcome the limitations of evaluations based on single-run results, repeated training was performed 10 times under a fixed baseline condition while maintaining the same data split and changing only the random seed. In addition, the effects of dataset size and hyperparameters were examined separately from the repeated-training stability evaluation through sensitivity analysis. The results showed that EfficientNet achieved the highest average performance and the lowest variability, with a Test Macro F1-score of 0.9566±0.0022. MobileNet exhibited a favorable balance between classification performance and computational efficiency based on its high performance and small model size. ResNet showed stable performance but had the largest number of parameters, whereas GoogLeNet exhibited low average performance and high variability under the present analysis conditions. Increasing the dataset size positively affected performance improvement and variability reduction for EfficientNet, MobileNet, and ResNet; however, the same trend was not observed for GoogLeNet. The OFAT analysis showed that the learning rate was a major influencing factor. These results indicate that CNN model selection for concrete damage classification should consider not only the highest performance from a single run but also repeated-training stability and computational efficiency.
|