|
| Ref. | Training function | Number of images | Image size | Training/testing ratio | Influence of training set size | Different types of cracks | Values of quality assessment criteria | Application |
|
| Xu et al. [25] | Momentum optimization algorithm | 6,069 | 512 × 512 | 66/34 | No | No | Accuracy = 96.37%, precision = 78.11%, recall = 100% | Crack detection in bridge infrastructures |
| Specificity = 95.83%, F1-score = 87.71% |
| Dung and Anh [28] | — | 40,000 | 227 × 227 | 80/10/10 | No | No | Accuracy training = 91.9%, accuracy validation = 89.6%, accuracy testing = 89.3% | Crack on the concrete surface in buildings |
| Chen et al. [23] | Adam algorithm | 40,000 | 227 × 227 | — | No | No | Accuracy = 99.71% | Crack on the concrete surface in buildings |
| Nahvi and Jabbari[9] | Stochastic gradient descent | 400 | 150 × 150 | — | No | No | Accuracy = 92.08%, precision = 100%, recall = 83% | Crack detection in the pavement surface |
| Zhang et al. [19] | Gradient descent | 3500 | 256 × 256 | 80/20 | Yes | No | Accuracy = 92.27% | Crack on the concrete surface after mechanical testing |
| This work | Stochastic gradient descent with momentum | 40,000 | 227 × 227 | 50/50 | No | No | Accuracy = 97.7%, precision = 96.5%, recall = 98.8% | Crack on the concrete surface in buildings |
| Specificity = 96.6%, F1-score = 97.7% |
|
