Table 1.
Comparison of existing recognition methods and the new method proposed by this study (d’ means the d-prime value of Equation (4). EER means equal error rate, and its concept is explained in Section 3.4). LBP, local binary patterns; RBWT, reverse biorthogonal wavelet transform; WCPH, weighted co-occurrence phase histograms; CLAHE, contrast-limited adaptive histogram equalization.
| Category | Method | Periocular Region | Accuracy | Advantage | Dis-Advantage |
|---|---|---|---|---|---|
| NIR camera- based | Personalized weight map [14] | Not using | EER of 0.78% (A) | Better image quality and recognition performance than the visible-light camera method | - Large and expensive NIR illuminator with NIR camera
- Additional power usage by NIR illuminator |
| SVM with hamming distance (HD) [15] | EER of 0.09% (B) | ||||
| EER of 0.12% (C) | |||||
| Fusion (AND rule) of left and right irises [16] | Accurate EER is not reported (EER of about 18–21% (D)) | ||||
| Adaptive bit shifting for matching by in-plane rotation angles [17] | EER of 4.30% (D) | ||||
| LBP with iris and periocular image in polar coordinate [21] | Using | EER of 10.02% (D) | |||
| Log-Gabor binary features with geometric key encoded features [41] | Not using | EER of 19.87%, d’ of 1.72 (E) |
Same algorithm for NIR and visible light iris images | Using manual hand-crafted features | |
| EER of 3.56%, d’ of 5.32 (D) | |||||
| CNN-based method (Proposed method) |
Using | EER of 3.04–3.10% (D) | Intensive CNN training is necessary | ||
| Visible light camera- based | Log-Gabor binary features with geometric key encoded features [41] | Not using | EER of 16.67%, d’ of 2.08 (F) |
Using manual hand-crafted features | |
| RBWT [29] | d’ of 1.09 (G) | Recognition is possible with general visible-light camera without additional NIR illuminator | - Image brightness is affected by environ- mental light
- Greater ghost effect caused by reflected light from environ- mental objects |
||
| Non-circular iris detection based on RANSAC [30] | d’ of 1.32 (G) | ||||
| Fusion of LBP and BLOBs features [31] | d’ of 1.48 (G) | ||||
| WCPH-based representation of local texture pattern [32] | d’ of 1.58 (G) | ||||
| CLAHE-based image enhancement [34] | EER of 18.82% (G) | ||||
| Pre-classification based on eyes and color [35] | EER of 16.94%, d’ of 1.64 (G) |
||||
| LBP-based periocular recognition [36] | Using | EER of 18.48%, d’ of 1.74 (G) |
|||
| AdaBoost training by multi-orient 2D Gabor feature [37] | Not using | d’ of 2.28 (G) | |||
| Combining color and shape descriptors [40] | EER of about 16%, d’ of about 2.42 (G) |
||||
| CNN-based method (Proposed method) |
Using | EER of 10.36%, d’ of 2.62 (G) |
Same algorithm for NIR and visible light iris images | Intensive CNN training is necessary | |
| EER of 16.25–17.9%, d’ of 1.87–2.26 (H) |
A: Institute of automation of Chinese academy of sciences (CAISA)-IrisV3-Lamp database; B: CASIA-Iris-Ver.1 database; C: Chek database; D: CASIA-Iris-distance database; E: Face recognition grand challenge (FRGC) database; F: University of Beira iris (UBIRIS).v2 database; G: NICE.II training dataset; H: Mobile iris challenge evaluation (MICHE) database.