| Title |
A Comparative Study on the Performance of Transfer Learning-Based CNN Backbone Models for Analyzing Reinforced Concrete Damage Images |
| DOI |
https://doi.org/10.12652/Ksce.2025.45.4.0513 |
| Keywords |
철근콘크리트; 손상 가지; 컨볼루션 신경망; 전이학습; 정확도; F1-score; 혼동행렬 Reinforced concrete; Damage detection; Convolution neural network(CNN); Transfer learning; Accuracy; F1-score; Confusion matrix |
| Abstract |
This study aims to effectively analyze damage images of reinforced concrete structures by comparing the performance of various convolutional neural network backbone models based on transfer learning, in order to identify the optimal architecture. A total of 3,500 damage images were used in experiments involving 12 pretrained models and 16 combinations of hyperparameters. The results showedthat while the highest top-1 validation accuracy of a vanilla CNN was limited to 67.5 %, the accuracy significantly improved to 86.0 % when using the EfficientNetB7 model with transfer learning, clearly demonstrating the benefit of transfer learning in this domain. However, classification performance cannot be fully evaluated by accuracy alone. In terms of F1-score, the InceptionV3 model exhibited the most balanced performance. Although the MobileNet series showed excellent efficiency and lightweight characteristics, it had limitations in precise classification under class-imbalanced conditions. Since high-precision classification is essential for structural damage image analysis, this study confirms the importance of considering comprehensive performance metrics such as F1-score in addition to accuracy. Furthermore, the findings suggest that additional validation under diverse environmental conditions and improvements in real-time processing capabilities are necessary for practical applications. Future work should focus on addressing data imbalance, enhancing the performance of lightweight models, and developing optimization techniques. |