Вернуться к статье
Обнаружение фальсифицированных подписей с применением глубинного обучения
Results on GPDS Synthetic
Method | Model | Accuracy, % | EER, % |
Euclidean distance | SigNet | 77.76 | – |
SigCNN | – | 7.34 | |
MSDN | – | 8.24 | |
TransOSV | – | 10.6 | |
HTCSigNet | 95.7 | 4.6 | |
ConvNeXt-B | 69.0 | 30.9 | |
ConvNeXt-L | 75.68 | 22.7 | |
CoAtNet-B | 68.0 | 31.5 | |
CoAtNet-L | 72.0 | 32.5 | |
Siamese ConvNeXt-B | 79.2 | 20.6 | |
Siamese CoAtNet-B | 79.5 | 20.2 | |
Cosine similarity | ConvNeXt-B | 70.7 | 29.0 |
ConvNeXt-L | 70.4 | 29.1 | |
CoAtNet-B | 69.4 | 29.6 | |
CoAtNet-L | 71.4 | 28.3 | |
Siamese ConvNeXt-B | 80.0 | 19.9 | |
Siamese CoAtNet-B | 78.9 | 20.9 |