Table 6.
Comparison of recognition accuracy among the previous method, various CNN nets, and the proposed method.
| Method Name | Input Image | Features (or Values) Used for Recognition | EER (%) | ||
|---|---|---|---|---|---|
| Good-Quality Database | Mid-Quality Database | Low-Quality Database | |||
| Previous method [12] | Original image | - | 0.474 | 2.393 | 8.096 |
| A (VGG Face (no fine-tuning)) | Fc7 | 1.536 | 5.177 | 7.264 | |
| A-1 (VGG Face (fine-tuning)) | Fc7 | 0.858 | 3.214 | 7.044 | |
| B (VGG Net-16 (no fine-tuning)) | Fc7 | 1.481 | 4.928 | 7.278 | |
| B-1 (VGG Net-16 (fine-tuning)) | Fc7 | 0.804 | 2.967 | 6.115 | |
| C (VGG Net-19 (no fine-tuning)) | Fc7 | 4.001 | 8.216 | 6.692 | |
| C-1 (VGG Net-19 (fine-tuning)) | Fc7 | 1.061 | 6.172 | 6.443 | |
| D (Revised Alexnet-1 (whole training)) | Difference image | Fc8 | 0.901 | 8.436 | 8.727 |
| E (Revised Alexnet-2 (whole training)) | Fc8 | 0.763 | 4.767 | 6.540 | |
| F (VGG Net-16 (fine-tuning) (proposed method)) | Fc8 | 0.396 | 1.275 | 3.906 | |