Automatic detection of cracks in structures using smartphone images

Early and accurate detection of cracks in structures is crucial for planning structural health and maintenance processes. This study aims to automatically detect cracks on wall and bridge surfaces using deep learning methods with smartphone images. The wall and bridge datasets consist of "cracked" and "crack-free" classes. Eighteen architectures, including convolutional neural networks, transformers, and hybrid architectures, were compared on the local dataset and the SDNET2018 dataset in terms of accuracy, positive predictive value, sensitivity, specificity, F1 score, AUC, and inference time per image. In local experiments, the proposed model achieved 98.50% accuracy on the wall dataset and 100.00% accuracy on the bridge dataset. In contrast, on zero-sample external datasets, accuracy dropped to 52.43%–73.53%. After raw co-training, the SDNET2018 dataset achieved accuracy rates of 86.05%, 81.35%, and 87.21% in wall, bridge deck, and road pavement tests, respectively; in balanced co-training, these values were 82.84%, 80.86%, and 86.57%. The findings show that data-source differences are decisive for model generalization and that multi-domain co-training can improve performance on external datasets. In conclusion, crack detection based on smartphone images can serve as a rapid, low-cost preliminary assessment approach that supports expert inspection in disasters and structural health management.