Deep Convolutional Neural Network Based on Two-Stream Convolutional Unit

Congcong Hou; Yuqing He; Xiaoheng Jiang; Jing Pan

doi:10.3788/LOP55.021005

Journals >Laser & Optoelectronics Progress >Volume 55 >Issue 2 >Page 021005 > Article

Laser & Optoelectronics Progress
Vol. 55, Issue 2, 021005 (2018)

Deep Convolutional Neural Network Based on Two-Stream Convolutional Unit

Congcong Hou^1、*, Yuqing He¹, Xiaoheng Jiang¹, and Jing Pan¹

Author Affiliations

¹ School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China

¹ School of Electronic Engineering, Tianjin University of Technology and Education, Tianjin 300222, China

show less

DOI: 10.3788/LOP55.021005 Cite this Article Set citation alerts

Congcong Hou, Yuqing He, Xiaoheng Jiang, Jing Pan. Deep Convolutional Neural Network Based on Two-Stream Convolutional Unit[J]. Laser & Optoelectronics Progress, 2018, 55(2): 021005 Copy Citation Text

show less

Fig. 1. Extraction process of features containing information in channels (a) and across the channels (b)

Download full size

Fig. 2. Simplified convolution

Download full size

Fig. 3. Comparison of two-stream convolutional unit (a) and simplified convolutional unit (b)

Download full size

Fig. 4. Features concat

Download full size

Fig. 5. Diagram of two-stream convolutional unit (a) and simplified convolutional unit (b)

Download full size

Fig. 6. Connection between convolutional units

Download full size

Fig. 7. Architecture of CTsNet for CIFAR database

Download full size

Layer	LSNet^[14]	CTsNet
1	3×3×3×192	3×3×3×192
2	$\begin{matrix} [\begin{matrix} 3 \times 3 \times 1 \times 192 \\ 1 \times 1 \times 192 \times 192 \end{matrix}] \end{matrix}$	$\begin{matrix} [\begin{matrix} (3 \times 3 \times 1 \times 192), (1 \times 1 \times 192 \times 192) \\ 1 \times 1 \times (192 \times 2) \times 192 \end{matrix}] \end{matrix}$
3	$\begin{matrix} [\begin{matrix} 3 \times 3 \times 1 \times 192 \\ 1 \times 1 \times 192 \times 192 \end{matrix}] \end{matrix}$	$\begin{matrix} [\begin{matrix} (3 \times 3 \times 1 \times 192), (1 \times 1 \times 192 \times 192) \\ 1 \times 1 \times (192 \times 2) \times 192 \end{matrix}] \end{matrix}$
4	$\begin{matrix} [\begin{matrix} 3 \times 3 \times 1 \times 192 \\ 1 \times 1 \times 192 \times 192 \end{matrix}] \end{matrix}$	$\begin{matrix} [\begin{matrix} (3 \times 3 \times 1 \times 192), (1 \times 1 \times 192 \times 192) \\ 1 \times 1 \times (192 \times 2) \times 192 \end{matrix}] \end{matrix}$
5	3×3×192×256	3×3×192×192
6	$\begin{matrix} [\begin{matrix} 3 \times 3 \times 1 \times 256 \\ 1 \times 1 \times 256 \times 256 \end{matrix}] \end{matrix}$	$\begin{matrix} [\begin{matrix} (3 \times 3 \times 1 \times 192), (1 \times 1 \times 192 \times 192) \\ 1 \times 1 \times (192 \times 2) \times 192 \end{matrix}] \end{matrix}$
7	$\begin{matrix} [\begin{matrix} 3 \times 3 \times 1 \times 256 \\ 1 \times 1 \times 256 \times 256 \end{matrix}] \end{matrix}$	$\begin{matrix} [\begin{matrix} (3 \times 3 \times 1 \times 192), (1 \times 1 \times 192 \times 192) \\ 1 \times 1 \times (192 \times 2) \times 192 \end{matrix}] \end{matrix}$
8	$\begin{matrix} [\begin{matrix} 3 \times 3 \times 1 \times 256 \\ 1 \times 1 \times 256 \times 256 \end{matrix}] \end{matrix}$	$\begin{matrix} [\begin{matrix} (3 \times 3 \times 1 \times 192), (1 \times 1 \times 192 \times 192) \\ 1 \times 1 \times (192 \times 2) \times 192 \end{matrix}] \end{matrix}$
9	3×3×256×256	3×3×192×192
10	3×3×256×256	3×3×192×192
11	1×1×256×10	1×1×192×10

Table 1. Configurations of network parameters

Method	Error /%	Parameter /M
LSNet^[14]	6.86	1.95
CTsNet	6.33	1.67

Table 2. Comparison on parameters and classification error rate of CTsNet and LSNet

Method	CIFAR10	CIFAR10+	CIFAR100
Maxout^[18]	11.68	9.38	38.57
NIN^[13]	10.41	8.81	35.68
NIN+LA^[23]	9.59	7.51	34.40
FitNet^[21]	-	8.39	35.04
DSN^[24]	9.75	8.22	34.57
Highway^[25]	-	7.54	32.24
ALL-CNN^[19]	9.08	7.25	33.71
RCNN-160^[26]	8.69	7.10	31.75
ResNet-110^[12]	-	6.43	-
CSNet-M^[20]	8.15	6.38	30.24
CTsNet	7.95	6.33	30.12

Table 3. Classification error rate of CIFAR10 and CIFAR100 with various networks%

Congcong Hou, Yuqing He, Xiaoheng Jiang, Jing Pan. Deep Convolutional Neural Network Based on Two-Stream Convolutional Unit[J]. Laser & Optoelectronics Progress, 2018, 55(2): 021005

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information