Table 5.
Comparison on the performance of different methods proposed in recent years.
| No | Studies | Year | Method | Parameters/FLOPs | Evaluation Protocol | Augmentation | Result |
|---|---|---|---|---|---|---|---|
| 1 | Othman et al. [50] | 2016 | IrisCode (2D-Gabor filter + Hamming Distance) | -/- | CASIA-V4: 602 classes (for testing) | None | CASIA-V4: 3.5% (Verification EER) |
| 2 | Nguyen et al. [43] | 2017 | Pre-trained CNN (Dense-Net) + SVM | -/- | CASIA-V4: 1000 classes (Train: 70%, Test: 30%) * | None | CASIA-V4: 98.8% (Identification Accuracy) |
| 3 | Alaslani et al. [67] | 2018 | Pre-trained CNN (Alex-Net) + SVM | 41 M/2.2 B | CASIA-V1: 60 classes CASIA-V3: 60 classes CASIA-V4: 60 classes (Train: 70%, Test: 30%) * |
None | CASIA-V1: 98.3% CASIA-V3: 89% CASIA-V4: 98% (Identification Accuracy) |
| 4 | Wang et al. [63] | 2018 | MiCoRe-Net | >1.4 M/>50 M | CASIA-V3: 218 classes (Train: 1346 images, Test: 218 images) * CASIA-V4: 1000 classes (Train: 9000 images, Test: 1000 images) * |
Rotation and Cropping | CASIA-V3: 99.08% CASIA-V4: 88.7% (Identification Accuracy) |
| 5 | Tobji et al. [64] | 2019 | FMnet | 15 K/10 M | CASIA-V4: 1000 classes (Train: 70%, Test: 30%) * | None | CASIA-V4: 95.63% (Identification Accuracy) |
| 7 | Boyd et al. [68] | 2019 | Pre-trained/Finetuned CNN (ResNet-50) + SVM | 25 M/5.1 B | CASIA-V4: 1000 classes (Train: 70%, Test: 30%) * | None | CASIA-V4: 99.03% (Identification Accuracy) |
| 6 | Liu et al. [45] | 2019 | Fuzzified image + Capsule network | >4 M/- | CASIA-V4: 1000 classes (Train: 80%, Test: 20%) | None | CASIA-V4: 83.1% (Identification Accuracy) |
| 8 | Lee et al. [65] | 2019 | Deep ResNet-152 +Matching distance | >53 M/>10 B | CASIA-V4: 1000 classes (Train: 50%, Test: 50%) | Translation and Cropping | CASIA-V4: 1.33% (Verification EER) |
| 9 | Proença et al. [49] | 2019 | VGG-19 based CNN | 138 M/- | CASIA-V4: 2000 classes (Train: 1000 classes, Test: 1000 classes) | Scale transform and Intensity transform | CASIA-V4: 3.0% (Verification EER) |
| 10 | Chen et al. [66] | 2020 | Tiny-VGG based CNN | >10 M/>1.3 B | CASIA-V4: 140 K pairs (Train: 50,632 images on another database) |
Contrast, Brightness, and Distortion | CASIA-V4: 99.58% (Identification Accuracy) CASIA-V4: 2.36% (Verification EER) |
| 11 | Proposed Method | 2021 | Condensed 2-ch CNN | 33 K/49.1 M | CASIA-V1: 108 classes (Finetune: 20 classes, Test: 88 classes) CASIA-V3: 233 classes (Train: 33 classes, Test: 200 classes) CASIA-V4: 648 classes (Finetune: 30 classes, Test: 615 classes) |
Brightness jitter, Horizontal shift, and Longitudinal scaling (Online) |
CASIA-V1: 100% CASIA-V3: 100% CASIA-V4: 99.77% (Identification Accuracy) CASIA-V1: 0.33% CASIA-V3: 0.76% CASIA-V4: 1.19% (Verification EER) |
* Training set and testing set share same classes. K-Kilo, M-Million, B-Billion.