Early Crack Detection in Civil Infrastructure Using Multi-Threshold Evaluation of Fine-Tuned ResNet Backbones on SDNET2018

Abstract

Early detection of cracks on civil infrastructure is vital for implementing preventive maintenance measures. In this paper, five different ImageNet pretrained models of the residual architecture ResNet-18, ResNet-34, ResNet-50, ResNet-101, and ResNetRS-101 are benchmarked for their performance in classifying crack images within the SDNET2018 dataset. Each model is trained using two different training procedures, namely training with a frozen backbone (NOFT) and training with end-to-end fine-tuning (FT). Due to the need for false positive and false negative rates to be considered in addition to the classification accuracy, a multi-threshold analysis for both metrics was conducted. The results of this study indicate that fine-tuning improves the classification performance of all five model backbones. Furthermore, the fine-tuned ResNet-34 model exhibits the best overall performance. At a threshold of 0.60 on the fixed test set of the SDNET2018 dataset (n = 8414), the model achieves an accuracy of 96.39%, precision of 87.58%, recall of 88.68%, and an F1-score of 88.13%, in addition to area under the ROC curve and area under the precision-recall curve values of 0.9805 and 0.9411, respectively. Furthermore, the model indicates 160 false positives and 144 false negatives at this threshold. Overall, then, the fine-tuned ResNet-34 model is selected as an efficient model to be utilized as a classification gate for early crack detection prior to performing crack segmentation.