Convolutional Neural Network Based on DenseNet Evolution for Image Classification Algorithm

Yongjie Ma; Peipei Liu

doi:10.3788/LOP57.241001

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 24 >Page 241001 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 24, 241001 (2020)

Convolutional Neural Network Based on DenseNet Evolution for Image Classification Algorithm

Yongjie Ma^* and Peipei Liu

Author Affiliations

College of Physics and Electronic Engineering, Northwest Normal University, Lanzhou, Gansu 730070, China

show less

DOI: 10.3788/LOP57.241001 Cite this Article Set citation alerts

Yongjie Ma, Peipei Liu. Convolutional Neural Network Based on DenseNet Evolution for Image Classification Algorithm[J]. Laser & Optoelectronics Progress, 2020, 57(24): 241001 Copy Citation Text

show less

Fig. 1. Structure diagram of CNN

Download full size

Fig. 2. Structure diagram of DenseNet

Download full size

Fig. 3. Flow chart of EA

Download full size

Fig. 4. Flow chart of D-ECNN algorithm

Download full size

Fig. 5. Partial images of the data set. (a) Positive sample; (b) negative sample

Download full size

Fig. 6. Verification accuracies of the two algorithms

Download full size

Fig. 7. Test accuracies of D-ECNN algorithm after 20 experiments

Download full size

Term	Parameter setting
Learning rate	[0.001,0.1]
Dropout rate	[0,0.1]
Number of filters in 2D convolution	[4,6,18]
Filter size for 2D convolution	[2,3]
Number of units	[64,128,256]
Layer or block	[2D convolution, fully connected layers, DenseNet]
Activation function of 2D convolutional layer	[Leaky ReLU, RelU, PReLU, ReLU]
Activation function of the fully connected layer	[Sigmoid, Softmax, ReLU]
Activation function of the last fully connected layer	Sigmoid

Table 1. Parameter setting of experiment

Layer	D-ECNN	Output channel
2D convolutional layer	Conv(3×3)	64
2D convolutional layer	Conv(2×2)	16
Dense block	Conv(1×1)Conv(3×3)	8
	Conv(1×1)Conv(3×3)	8
	Conv(1×1)Conv(3×3)	8
Transition layer	Conv(1×1)average pool(2×2)	4
Fully connected layer	Sigmoid	64
Fully connected layer	Sigmoid	2

Table 2. Structure of D-ECNN model

Rate	Accuracy	Recall	Precision	F1-score
9∶1	95.78	95.50	71.73	81.92
8∶2	95.15	94.50	84.94	89.48
7∶3	95.36	94.31	90.62	92.43
6∶4	95.39	94.75	93.78	94.26
Average	95.42	94.77	85.28	89.52

Table 3. Test performances of D-ECNN model under different segmentation rates unit: %

Algorithm	Accuracy	Recall	Precision	F1-score
D-ECNN	95.34	94.88	95.77	95.32
VGG16	94.56	92.63	96.36	94.46

Table 4. Performance indicators of D-ECNN and VGG16 algorithms unit: %

Algorithm	Number of network layers	Number of network parameters	Number of training parameters	Model file size /M	Time-consuming of the test data set /s
D-ECNN	11	70939	69703	1.04	0.0345
VGG16	16	27844930	27834434	212	0.3967

Table 5. Train parameters of the two algorithms in the same train set

Algorithm model	Number of layers	Learning rate	Accuracy /%
D-ECNN	11	0.0050	95.34
Model-1	18	0.0035	94.53
Model-2	19	0.0037	93.42
Model-3	13	0.0059	94.22
Model-4	19	0.0047	90.73
Model-5	11	0.0045	68.76
Model-6	19	0.0055	94.82
Model-7	5	0.0058	94.38
Model-8	18	0.0055	94.83
Model-9	6	0.0045	95.23

Table 6. Parameters of 10 algorithm models

Yongjie Ma, Peipei Liu. Convolutional Neural Network Based on DenseNet Evolution for Image Classification Algorithm[J]. Laser & Optoelectronics Progress, 2020, 57(24): 241001

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information