Evaluating Transfer-Learning Strategies for Multi-Label Classification of Concrete Bridge Defects with Limited Data

Abstract

Bridge inspection images often contain multiple co-occurring surface defects within the same scene, yet much of the existing literature has emphasized single-label classification tasks. This study examines how transfer learning strategy affects multi-label classification of concrete bridge defects using the Concrete Defect Bridge Image (CODEBRIM) dataset and three ImageNet-pretrained convolutional neural network backbones: VGG16, ResNet50, and EfficientNetB0. Each backbone was evaluated under an add-on configuration with a frozen backbone and a selective fine-tuning configuration, both using a common multi-label prediction head with six sigmoid outputs corresponding to the defect classes and the Background class. To further examine the effect of backbone adaptation, performance sensitivity to fine-tuning depth was analyzed by progressively varying the number of trainable final backbone layers. The results showed that selective fine-tuning consistently outperformed the frozen-backbone add-on configuration across all three backbones, and that performance did not improve simply by unfreezing more layers. Instead, moderate selective fine-tuning provided the strongest overall results. Among the evaluated configurations, VGG16 with the last three backbone layers unfrozen achieved the best holdout test-set performance, with a subset accuracy of 0.840 and an F1-score of 0.874. These findings indicate that transfer-learning design choices, particularly the extent of backbone adaptation, play a central role in multi-label bridge-defect classification on limited engineering datasets.