Image sentiment classification via deep learning structure optimization

	Global average pooling		Output size of dropout	Output size of full connection	Output size of Softmax
	Kernel size	Output size	Output size of dropout	Output size of full connection	Output size of Softmax
Auxiliary classifier of Layer (4a)	14×14	1×1×512	1×1×512	1×1×2	1×1×2
Auxiliary classifier of Layer (4b)	14×14	1×1×512	1×1×512	1×1×2	1×1×2
Auxiliary classifier of Layer (4d)	14×14	1×1×528	1×1×528	1×1×2	1×1×2

Table 2.
Details of the Twitter dataset
Twitter数据集详细信息
View table
View in Article
Table 2.
Details of the Twitter dataset
Twitter数据集详细信息
Five agree Four agree Three agree
Positive 581 689 769
Negative 301 427 500
Total 882 1 116 1 269

Table 3.
Results of comparative experiment
对比实验结果

View table

View in Article

Table 3.
Results of comparative experiment
对比实验结果

		PCNN^[25]	SentiNet-A^[3]	Proposed algorithm
注：加粗字体为每行最优值。
Five agree	P	0.770	0.895	0.900
	R	0.878	0.878	0.922
	F1	0.821	0.886	0.911
	A	0.747	0.851	0.881
Four agree	P	0.733	0.851	0.860
	R	0.845	0.835	0.877
	F1	0.785	0.842	0.868
	A	0.714	0.807	0.836
Three agree	P	0.714	0.823	0.836
	R	0.806	0.809	0.848
	F1	0.757	0.814	0.842
	A	0.687	0.777	0.807

Table 4.
Comparative results of adding auxiliary classifiers at each layer
各层添加辅助分类器作用对比
View table
View in Article
Table 4.
Comparative results of adding auxiliary classifiers at each layer
各层添加辅助分类器作用对比
Number Convolution layer of auxiliary classifier Accuracy
(1) No auxiliary classifier 0.800
(2) Layer (3a) 0.817
(3) Layer (3b) 0.808
(4) Layer (4c) 0.798
(5) Layer (4e) 0.815

Table 5.
Ablation experiment results
消融实验结果对比

View table

View in Article

Table 5.
Ablation experiment results
消融实验结果对比

Number	Ablation network structure	Accuracy
(1)	Our algorithm	0.881
(2)	Delete the auxiliary classifier of layer (4a)	0.862
(3)	Delete the auxiliary classifier of Layer (4b)	0.866
(4)	Delete the auxiliary classifier of Layer (4d)	0.858
(5)	Delete the auxiliary classifier of Layer (4a) and Layer (4b)	0.849
(6)	Delete the auxiliary classifier of Layer (4a) and Layer (4d)	0.845
(7)	Delete the auxiliary classifier of Layer (4b) and Layer (4d)	0.851

Tools

Get Citation

Copy Citation Text

Jiachuan Sheng, Yaqi Chen, Jun Wang, Yahong Han. Image sentiment classification via deep learning structure optimization[J]. Infrared and Laser Engineering, 2020, 49(11): 20200269

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites