Expression Recognition Based on Attention-Split Convolutional Residual Network

Jiamin Chen; Yang Xu

doi:10.3788/LOP202259.1815009

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 18 >Page 1815009 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 18, 1815009 (2022)

Expression Recognition Based on Attention-Split Convolutional Residual Network

Jiamin Chen¹ and Yang Xu^1、2、*

Author Affiliations

¹College of Big Data and Information Engineering , Guizhou University, Guiyang 550025, Guizhou , China

²Guiyang Aluminum-Magnesium Design and Research Institute Co. Ltd., Guiyang 550009, Guizhou , China

show less

DOI: 10.3788/LOP202259.1815009 Cite this Article Set citation alerts

Jiamin Chen, Yang Xu. Expression Recognition Based on Attention-Split Convolutional Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815009 Copy Citation Text

show less

Fig. 1. Convolution. (a) Group convolution; (b) point convolution

Download full size

Fig. 2. Split convolution

Download full size

Fig. 3. Coordinate attention

Download full size

Fig. 4. CASCBlock

Download full size

Fig. 5. Channel attention resolution convolutional residuals network

Download full size

Fig. 6. Results of FER2013 experiment. (a) ResNet18; (b) SPResNet18_CA

Download full size

Fig. 7. Results of RAF-DB experiment. (a) ResNet18; (b) SPResNet18_CA

Download full size

Fig. 8. Accuracy curve comparison chart. (a) FER2013; (b) RAF-DB

Download full size

Layer	Output	Proposed network
	44×44	7×7，64，s=1
	44×44	3×3 maxpool，s=1
1	44×44	$[\begin{array}{l} S P C o n v, 64 \\ S P C o n v, 64 \\ C A \end{array}] \times 2$
2	22×22	$[\begin{array}{l} S P C o n v, 128 \\ S P C o n v, 128 \\ C A \end{array}] \times 2$
3	11×11	$[\begin{array}{l} S P C o n v, 256 \\ S P C o n v, 256 \\ C A \end{array}] \times 2$
4	6×6	$[\begin{array}{l} S P C o n v, 512 \\ S P C o n v, 512 \\ C A \end{array}] \times 2$
	1×1	Global avgpool 7‑d fc

Table 1. Parameters of channel attention resolution convolutional residuals network

Model name	Number of parameters	Model size /Mbit	Accuracy in FER2013 /%	Accuracy in RAF-DB /%
ResNet18	11180103	42.7	69.769	81.845
SPResNet18	4188487	16.2	72.444	83.638
SPResNet18_CA	4263223	16.5	72.666	84.420

Table 2. Model performance

Data	NetWork	Accuracy /%
FER2013	SHCNN^［1］	69.1
	LCMA^［21］	69.6
	LIANG^［22］	70.3
	Zhou^［23］	70.9
	SPResNet18_CA	72.6

Table 3. Comparative experiment on FER2013

Data	NetWork	Accuracy /%
RAF-DB	DLP-CNN^［24］	74.20
	EAU-Net^［25］	81.83
	DeepExp3D^［26］	82.06
	pACNN^［27］	83.05
	SPResNet18_CA	84.42

Table 4. Comparative experiment on RAF-DB

Jiamin Chen, Yang Xu. Expression Recognition Based on Attention-Split Convolutional Residual Network[J]. Laser & Optoelectronics Progress, 2022, 59(18): 1815009

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information