Face Recognition Based on Lightweight Recursive Residual Neural Network

Input	Operator	E	C	n	s
112×112×3	Conv 3×3	-	64	1	2
56×56×64	DW Conv 3×3	-	64	1	1
56×56×64	Bottleneck	3/2	64	5	2
28×28×64	Bottleneck	3	128	1	2
14×14×128	Bottleneck	3/2	128	6	1
14×14×128	Bottleneck	3	128	1	2
7×7×128	Bottleneck	3/2	128	2	1
7×7×128	Conv 1×1	-	512	1	1
7×7×512	Linear WGAP	-	512	1	1
1×1×512	Linear Conv 1×1	-	128	1	1

Table 2. Different model accuracy and parameter ratio

View table

Table 2. Different model accuracy and parameter ratio

Network	LFW/%	AgeDB/%	CFP-FP/%	FLOP	Speed/ms	Model size/MB	Parameter number /10⁶
ShiftFaceNet^[16]	96.00	-	-	-	-	3.1	0.780
Light CNN-29^[17]	99.33	-	-	-	-	50.0	4.000
LMobileNetE^[13]	99.25	95.12	90.68	3.17×10⁸	24	112.0	3.200
MobileFacenet^[9]	99.29	95.58	91.17	4.33×10⁸	39	4.0	0.990
M1	99.37	95.98	91.60	3.84×10⁸	34	3.9	0.818
M2	99.20	94.68	90.42	3.77×10⁸	31	3.9	0.795
M4	99.37	96.08	92.02	3.84×10⁸	34	3.9	0.816
M5	99.22	94.72	90.53	3.77×10⁸	31	3.9	0.793

Table 3. Comparison of model recognition accuracy unit: %
View table
Table 3. Comparison of model recognition accuracy unit: %
Network LFW AgeDB CFP-FP
SE-LResNet18^[13] 99.53 96.18 93.42
SE-LM6-18 99.62 96.95 94.34
SE-LM7-18 99.67 96.92 94.23

Tools

Get Citation

Copy Citation Text

Xiuling Zhang, Kaixuan Zhou, Qijun Wei, Jinxiang Li. Face Recognition Based on Lightweight Recursive Residual Neural Network[J]. Laser & Optoelectronics Progress, 2020, 57(22): 221506

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites