Table 3.
The average test classification accuracy.
DCNN model | Augmentation | Fine-tuning | LR | Accuracy (%) |
---|---|---|---|---|
VGG-16 | 5e-6 | 38.96 | ||
✓ | 5e-6 | 56.76 | ||
✓ | 5e-6 | 91.15 | ||
✓ | ✓ | 5e-6 | 97.19 | |
ResNet-50 | 5e-3 | 57.47 | ||
✓ | 1e-3 | 57.74 | ||
✓ | 5e-3 | 93.45 | ||
✓ | ✓ | 7.5e-3 | 96.86 | |
SqueezeNet | 3e-5 | 47.42 | ||
✓ | 3e-5 | 62.81 | ||
✓ | 3e-5 | 78.58 | ||
✓ | ✓ | 3e-5 | 90.71 |
Each model is trained with four settings of dynamic data-augmentation and fine-tuning. When a model is not fine-tuned, it is trained from scratch with random initialization.