Table 2.
The ablation study under different model configuration.
No. | Structure | Loss | Reg. | Aug. | Att. | Pru. | Fin. | EER |
---|---|---|---|---|---|---|---|---|
1 | Resnet | MSE | √ | √ | × | × | × | 1.55% |
2 | A | MSE | √ | √ | × | × | × | 1.26% |
3 | A | MSE | × | √ | × | × | × | 1.86% |
4 | A | MSE | √ | × | × | × | × | 3.40% |
5 | A | CE | √ | √ | × | × | × | 1.55% |
6 | A | MSE | √ | √ | √ | × | × | 1.09% |
7 | C | MSE | √ | √ | √ | √ | × | 1.22% |
8 | C | MSE | √ | √ | √ | √ | √ | 1.03% |
9 | C † | MSE | √ | √ | √ | √ | √ | 1.26% |
10 | C o | MSE | √ | √ | √ | √ | √ | 0.76% |
Reg., Aug., Att., Pru., and Fin. are the abbreviation of regularization term, augmentation layer, attention layer, pruning, and finetuning. † The model is trained from scratch. o The uncropped iris images with resolution of 30 × 360 serve as input.