1 Introduction
2 Background
2.1 Curvelet transform
2.2 Fractal dimension
2.3 Deep learning approaches
3 The proposed framework
3.1 The proposed Curvelet–Fractal approach
3.1.1 Curvelet via wrapping transform
3.1.2 Fractional Brownian motion method
3.1.3 Improved differential box counting method
3.1.4 Face matching
3.2 Learning additional features representations
4 Experimental results
4.1 Face datasets
-
SDUMLA-HMT dataset [48] This includes 106 subjects and each has 84 face images taken from 7 viewing angles and under different experimental conditions including, facial expressions, accessories, poses, and illumination. The main purpose of this dataset is to simulate real-world conditions during face image acquisition. The image size is \((\mathbf{640} \times \mathbf{480})\) pixel.
-
FERET dataset [49] This contains a total of 14,126 images taken from 1196 subjects, with at least 365 duplicate sets of images. This is one of the largest publicly available face datasets with a high degree of diversity of facial expression, gender, illumination conditions and age. The image size is \((\mathbf{256} \times \mathbf{384})\) pixel.
-
CAS-PEAL-R1 dataset [50] A subset of the CAS-PEAL face dataset has been released for research purposes and named CAS-PEAL-R1. This contains a total of 30,863 images taken from 1040 Chinese subjects (595 are males and 445 are females). The image size is \((\mathbf{360} \times \mathbf{480})\) pixel.
-
LFW dataset [51] This contains a total of 13,233 images taken from 5749 subjects where 1680 subjects appear in two or more images. In the LFW dataset, all images were collected from Yahoo! News articles on the Web, with a high degree of intra-personal variations in facial expression, illumination conditions, occlusion from wearing hats and glasses, etc. It has been used to address the problem of unconstrained face verification task in recent years. The image size is \((\mathbf{250} \times \mathbf{250})\) pixel.
4.2 Face identification experiments
4.2.1 Parameter settings of the Curvelet–Fractal approach
4.2.2 MDFR architecture and training details
DBN models | Accuracy rate % |
---|---|
600–600–1000 | 92.19 |
700–700–1000 | 92.48 |
800–800–1000 | 95.38 |
900–900–1000 | 93.68 |
4.2.3 Comparative study of fractal, Curvelet–Fractal, DBN and MDFR approaches
Approach | Fb | Fc | Dup.I | Dup.II |
---|---|---|---|---|
DLBP-W [56] | 99 | 99 | 86 | 85 |
G-LQP [57] | 99.9 | 100 | 93.2 | 91.0 |
FHOGC [55] | 98.3 | 98.3 | 86.3 | 81.2 |
Groupwise MRF [54] | 98.5 | 98.8 | 87.7 | 86.2 |
H-Groupwise MRF [54] | 99.7 | 99.2 | 94.7 | 93.6 |
LGOP+WPCA [58] | 99.2 | 99.5 | 89.5 | 88.5 |
DFD(\(\hbox {S}=3\))+WPCA [53] | 99.3 | 99 | 88.8 | 87.6 |
DFD(\(\hbox {S}=5\))+WPCA [53] | 99.4 | 100 | 91.8 | 92.3 |
AMF [11] | 99.9 | 100 | 96.4 | 93.6 |
GOM [10] | 99.9 | 100 | 95.7 | 93.1 |
DBN | 99.91 | 100 | 95.15 | 93.35 |
\({Fractal}_{Vector}\)
| 97.5 | 96.65 | 92 | 90.34 |
Curvelet–Fractal
| 100 | 98.97 | 97.92 | 95.72 |
MDFR framework | 100 | 100 | 98.40 | 97.86 |
4.3 Face verification experiments
Approach | PE | PA | PL | PT | PB | PS |
---|---|---|---|---|---|---|
RBFNN [59] | 84.8 | 93.4 | 63.4 | 96.9 | – | – |
DT-LBP [60] | 98 | 92 | 41 | – | – | – |
DLBP-W [56] | 99 | 92 | 41 | – | – | – |
1D-CFA [61] | 83.12 | 74.84 | 31.43 | 71.21 | 98.19 | 98.55 |
Groupwise MRF [54] | 94.8 | 90.3 | 66.9 | 99.2 | 98.8 | 99.5 |
H-Groupwise MRF [54] | 96.4 | 90.3 | 66.9 | 99.8 | 100 | 99.6 |
LGOP+WPCA [58] | 99.6 | 96.8 | 69.9 | – | – | – |
DFD(\(S=3\))+WPCA [53] | 99 | 96.9 | 63.9 | – | – | – |
DFD(\(S=5\))+WPCA [53] | 99.6 | 96.9 | 58.9 | – | – | – |
FHOGC [55] | 94.9 | 90.3 | 68.7 | 100 | – | – |
LBP [62] | 92.93 | 82.58 | 32.46 | – | – | – |
DBN | 98.93 | 75.36 | 80.60 | 95.45 | 96.01 | 97.09 |
Fractal
\(_{Vector}\)
| 95.12 | 92.55 | 78.01 | 92.33 | 95.23 | 96.03 |
Curvelet–Fractal
| 99.87 | 98.07 | 89.48 | 100 | 100 | 99.64 |
MDFR Framework | 100 | 99.43 | 89.92 | 100 | 100 | 100 |
Approach | Acc. \(({\hat{\upmu }} \pm \hbox {S}_{\mathrm{E}})\)
| Protocol |
---|---|---|
DeepFace [67] | 0.9735 ± 0.0025 | Unrestricted |
DeepID [47] | 0.9745 ± 0.0026 | Unrestricted |
ConvNet-RBM[68] | 0.9252 ± 0.0038 | Unrestricted |
Convolutional DBN [18] | 0.8777 ± 0.0062 | Restricted |
DDML [64] | 0.9068 ± 0.0141 | Restricted |
VMRS [69] | 0.9110 ± 0.0059 | Restricted |
HPEN+HD-LBP+DDML [65] | 0.9257 ± 0.0036 | Restricted |
HPEN+HD-Gabor+DDML [65] | 0.9280 ± 0.0047 | Restricted |
Sub-SML+Hybrid+LFW3D [70] | 0.9165 ± 0.0104 | Restricted |
MSBSIF-SIEDA [66] | 0.9463 ± 0.0095 | Restricted |
DBN | 0.9353 ± 0.0165 | Restricted |
Curvelet–Fractal | 0.9622 ± 0.0272 | Restricted |
MDFR framework | 0.9883 ± 0.0121 | Restricted |
Database | DBN | Curvelet–Fractal | MDFR framework |
---|---|---|---|
SDUMLA-HMT | 18 h & 35 min | 35 min | 4 h & 15 min |
FERET | 16 h & 45 min | 17 min | 3 h & 33 min |
CAS-PEAL-R1 | 15 h & 27 min | 14 min | 3 h & 32 min |
LFW-a | 13 h & 41 min | 7 min | 2 h & 56 min |