Researching | Multi-Pose Face Recognition with Two-Cycle Generative Adversarial Network

Acta Optica Sinica, Volume. 40, Issue 19, 1910002(2020)

Multi-Pose Face Recognition with Two-Cycle Generative Adversarial Network

Zhijing Xu and Dong Wang^*

Author Affiliations

College of Information Engineering, Shanghai Maritime University, Shanghai 201306, China

show less

Abstract Get PDF(in Chinese)

Figures&Tables (8)

References (22)

Paper Information

Figures & Tables(8)

Fig. 1. Framework of TC-GAN

Download full size

Fig. 4. Face frontalization results on Multi-PIE

Download full size

Fig. 5. Face rotation results on Multi-PIE

Download full size

Table 1. Detailed structure of TC-GAN

Table 1. Detailed structure of TC-GAN

Encoder and generator			Decoder
Layer	Filter/stride	Output size	Layer	Filter/stride	Output size
						FC		6×6×320
Conv11		3×3/1		96×96×32		FConv52	3×3/1	6×6×160
Conv12		3×3/1		96×96×64		FConv51	3×3/1	6×6×256
Conv21		3×3/2		48×48×64		FConv43	3×3/2	12×12×256
Conv22		3×3/1		48×48×64		FConv42	3×3/1	12×12×128
Conv23		3×3/1		48×48×128		FConv41	3×3/1	12×12×192
Conv31		3×3/2		24×24×128		FConv33	3×3/2	24×24×192
Conv32		3×3/1		24×24×96		FConv32	3×3/1	24×24×96
Conv33		3×3/1		24×24×192		FConv31	3×3/1	24×24×128
Conv41		3×3/2		12×12×192		FConv23	3×3/2	48×48×128
Conv42		3×3/1		12×12×128		FConv22	3×3/1	48×48×64
Conv43		3×3/1		12×12×256		FConv21	3×3/1	48×48×64
Conv51		3×3/2		6×6×256		FConv13	3×3/2	96×96×64
Conv52		3×3/1		6×6×160		FConv12	3×3/1	96×96×32
Conv53		3×3/1		6×6×320		FConv11	3×3/1	96×96×3
AvgPool		6×6×1		1×1×320
FC(D_E/D_N)		N_d+N_p+1/ (N_d+1)

Table 2. Face recognition rates at different stagesunit:%
View table
Table 2. Face recognition rates at different stagesunit:%
Stage 0° ±15° ±30° ±45° ±60° ±75° ±90°
Stage-1 98.1 95.1 92.0 84.7 80.4 75.3 70.6
Stage-2 98.9 97.6 95.8 90.6 86.1 80.2 74.3

Table 3. Face recognition rates under different methodsunit:%

Table 3. Face recognition rates under different methodsunit:%

Method	0°	±15°	±30°	±45°	±60°	±75°	±90°	Avg1	Avg2
HPN^[21]	95.4	84.2	78.3	72.8	61.2	47.6	29.8	0.784	0.670
VGG-Face^[22]	98.1	97.2	96.1	92.6	84.7	62.8	34.2	0.937	0.808
TP-CNN	97.1	96.7	94.4	92.9	87.8	75.6	59.8	0.937	0.863
FF-GAN	95.7	94.6	92.5	89.7	85.2	77.2	61.2	0.915	0.852
DR-GAN^[23]	98.1	95.0	91.3	88.0	85.8	-	-	0.916	-
MTAN	96.5	95.3	89.7	87.9	84.1	-	-	0.907	-
TC-GAN	98.9	97.6	95.8	90.6	86.1	80.2	74.3	0.938	0.891

Table 4. Face recognition rates under different factorsunit:%
View table
Table 4. Face recognition rates under different factorsunit:%
Method 0° ±15° ±30° ±45° ±60° ±75° ±90°
After removing pixel-wise loss 99.7 95.3 90.2 85.1 84.4 77.2 72.9
After removing cycle loss 99.6 95.7 92.9 86.1 83.7 76.9 71.3
TC-GAN 98.9 97.6 95.8 90.6 86.4 80.2 74.3

Table 5. Recognition rate comparison on CFP database
View table
Table 5. Recognition rate comparison on CFP database
Method Frontal-Frontal Frontal-Profile
VGG-Face 97.59 89.12
TP-CNN 97.35 92.41
DR-GAN 97.08 91.02
Human 96.24 94.57
TC-GAN 97.41 93.13

Tools

Get Citation

Copy Citation Text

Zhijing Xu, Dong Wang. Multi-Pose Face Recognition with Two-Cycle Generative Adversarial Network[J]. Acta Optica Sinica, 2020, 40(19): 1910002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Image Processing

Received: May. 6, 2020

Accepted: Jun. 23, 2020

Published Online: Sep. 29, 2020

The Author Email: Wang Dong (2368072842@qq.com)

DOI:10.3788/AOS202040.1910002

Topics

laser devices and laser physics

Lasers and Laser Optics

laser manufacturing

Instrumentation, Measurement and Metrology