2020 IEEE 4th Information Technology,Networking,Electronic and Automation Control Conference (ITNEC 2020)

Research on Extended Image Data Set

Based on Deep Convolution Generative
Adversarial Network
Zixi Liu*, Ming Tong, Xiaoyu Liu, Zhixiong Du, Weicong Chen
[email protected]
Abstract—The Deep Convolution Generative Adversarial accuracy increased from 92.3% to 96.1% [2]. Wang gege
Network (DCGAN) adds the structure of Generative Adversarial pointed out that the linear convolution layer of DCGAN is a
Network (GAN) on the basis of the generation countermeasure generalized linear model for the lower data block, its abstract
network, and specially generates image samples. In this paper, level is low, and the quality of the generated image is not high,
DCGAN is used to generate the image which does not belong to and only the subjective visual perception is used to evaluate the
MNIST data set, and then, a new data set is obtained. Finally, image quality in the performance measurement of the model, a
Convolutional Neural Networks (CNN) [1] is used to test the new multilayer perceptron depth convolution generation
data set. We need define an initializer to make the GAN converge countermeasure network (MPDCGAN) is proposed, which is
better, and use the standard LEAKYRELU function to activate replaced by multilayer perceptron convolution layer. The
the GAN. The generator is defined by a fully connected layer with generalized linear model is convoluted on the input data to
an input size of 128. The noise is Gaussian white noise, which uses capture the deeper features of the image, and Frechet perception
RELU as the activation function. The discriminator uses two distance (FID) is used to measure the quality of image
convolution layers, the first one uses RELU as the activation generation [3]. Lu Ping and Dong Husheng designed an image
function, and the second layer uses sigmoid function. The results generation model based on the Deep Convolution Generative
show that the accuracy of the new data set is the same as that of Adversarial Network (DCGAN). The experimental results on
the original data set when tested on CNN, and the method of the CelebA data set show that the model has excellent effect on
expanding MNIST data set by using deep convolution is effective. face image generation [4].

II. GAN and DCGAN

Keywords: Deep Convolution Generative Adversarial Network
(DCGAN); MNIST; Generator; Discriminator Generative Adversarial Network [5] is a kind of neural
network, which has a generation network G (generator), which
I. INTRODUCE
is responsible for generating pictures. It receives a random noise
The full name of GAN is Generative Adversarial Network. Z, generates pictures through the noise, records the generated
The original GAN is a kind of unsupervised learning method. It pictures as 𝐺𝐺(𝑧𝑧) , and a discrimination network D
cleverly uses the idea of "against" to learn the generative model. (discriminator), which is responsible for judging whether a
Once the training is completed, it can generate new data samples. picture is "real". In the process of training, the goal of generating
DCGAN extends the concept of Gan to convolutional neural network G is to generate real pictures as much as possible to
network, which can generate high quality image samples. cheat the discrimination network D, and the goal of D is to
DCGAN is widely used in various fields. Master Mino uses the distinguish the generated pictures from the real pictures as much
Deep Convolution Generative Adversarial Network to learn the as possible. In this way, G and D constitute a dynamic "game",
distribution of micro Doppler radar image, and generates new which is the basic idea of GAN. Chen Lifang and Lu Guojun
image according to the learned distribution. The generated proposed a MNIST handwritten digital image generation
image is mixed with the original training set as a new training method based on VAE and Gan fusion network. This method
set to train the classification model. The experimental results combines the advantages of VAE and Gan, and has the accuracy
show that after adding the image generated by the depth of VAE image generation and the clarity of Gan image
convolution to the training set, the training model is tested. The generation. Experiments show that this method can generate

978-1-7281-4390-3/20/$31.00 ©2020 IEEE 47

Authorized licensed use limited to: University of Canberra. Downloaded on June 06,2020 at 16:15:18 UTC from IEEE Xplore. Restrictions apply.
 high quality handwritten digital image [6]. Zhang Jialiang variable of unknown distribution.
and others used the Variational Auto Encoder (VAE) to extract
According to the cross-entropy loss, the following loss
the real image information, and regarded the reconstructed
functions can be constructed:
image as a false image to improve the BEGAN [7]. Zhang
Guanghua and others first used GAN to generate astronomical V(D，G)=𝐸𝐸𝑥𝑥 ∼𝑃𝑃𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 (𝑥𝑥) [ln 𝐷𝐷(𝑥𝑥)]+ 𝐸𝐸𝑧𝑧 ∼𝑃𝑃𝑧𝑧 (𝑧𝑧) [ln(1 − 𝐷𝐷(𝐺𝐺(𝑧𝑧)))]（ 1 ）

images, and gave the GAN model structure of astronomical The specific steps of the experiment are as follows:
image generation [8].
Define the generator. The generator has two convolution
GAN does not have any structural restrictions on G and D, networks, and tanh is used as the activation function.
while Deep Convolutional Generative Adversarial Networks
 Define the discriminator. The input image of the
(DCGAN) [9] uses a special structure in D and G to effectively
discriminator is 28 × 28, 64 convolution kernels and 5 × 5
model pictures.
kernels. 64 5 × 5 convolution kernels are used to extract
III. EXPERIMENT different features, and the default step size is 1. There are four
In this paper, DCGAN is used to generate images in convolution networks. Tanh is used as the activation function of
TensorFlow. The data set is MNIST data set. Suppose that the the first, second and third layers, and sigmoid is used as the
real picture data used for training is x, and the distribution of activation function of the fourth layer.
picture data is 𝑃𝑃𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 (𝑥𝑥). Previously, it was said that G can  Define the training function. Select 128 for
"generate pictures", and the real G can learn the real data Batch_size, 100 for epoch, 0.0005 for learning rate, and SGD
distribution 𝑃𝑃𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 (𝑥𝑥) . The distribution of noise Z is set as random gradient descent algorithm for loss function.
𝑃𝑃𝑧𝑧 (𝑧𝑧), 𝑃𝑃𝑧𝑧 (𝑧𝑧) is known, and 𝑃𝑃𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 (𝑥𝑥) is unknown. In an ideal
Select part of the epoch output image experimental results,
case, the distribution of 𝐺𝐺(𝑧𝑧) should be as close as possible to
as follows:
𝑃𝑃𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 (𝑥𝑥). G maps the Z variable of known distribution to the X

Fig. 1. 0_epoch Fig. 2. 10_epoch

48

Authorized licensed use limited to: University of Canberra. Downloaded on June 06,2020 at 16:15:18 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. 40_epoch Fig. 4. 99_epoch

When the above four graphs are 0 epoch, 10 epoch, 40 distinguished.

epoch, 99 epoch, the depth convolution generates the new
The new data set is trained by convolution neural network,
MNIST image generated against the network. It can be seen that
which has three convolution layers. The selected activation
at 0 epoch, they are all noise, no image number. At the time of
function is ReLU, Batch_size is 64, epoch is 50, and the
10 epoch, there are already images. At this time, it is impossible
optimizer is Adadelta function. Select several rounds of
to distinguish what the image content is. At 40 epochs, there are
representative results as follows:
clear digital images, but some of them are still unclear. In the
last round of training, all the generated images can be
Table Ⅰ

Epoch Loss Acc

1 0.0313 0.9919

12 0.0373 0.9923

29 0.0422 0.9928

38 0.0274 0.9995

The trained model is saved and tested. The test result is Convolution Neural Network is used to test the new MNIST
0.9935. data set. The results show that the test accuracy of the
Convolution Neural Network to the new data set is equivalent
IV. SUMMARY
to that of the Convolution Neural Network on the original
In this paper, the concept and some applications of Deep MNIST data set, that is to say, the Deep Convolution Generative
Convolution Generative Adversarial Network and Generative Adversarial Network is proved. The new data generated can be
Adversarial Network are introduced. In this paper, the original used in practice. This experiment also proves that the Deep
MNIST data set is expanded to generate a new data set by using Convolution Generative Adversarial Network can use the
the Deep Convolution Generative Adversarial Network, and the

49

Authorized licensed use limited to: University of Canberra. Downloaded on June 06,2020 at 16:15:18 UTC from IEEE Xplore. Restrictions apply.
original data set to generate new data to meet the requirements
in the case of few data sets, which solves the problem of finding
data difficult and data less can not be effectively trained.

REFERENCES
[1] Waibel, A., 1987. Phoneme recognition using time-delay neural
networks. Meeting of the Institute of Electrical, Information and
Communication Engineers (IEICE). Tokyo, Japan.
[2] Miyo. Research on radar image enhancement technology based on
generated countermeasure network [D]. Beijing University of Posts and
telecommunications, 2019.
[3] Wang gege, Guo Tao, Li Guiyang. Multi layer perceptron deep
convolution generation countermeasure network [J]. Computer science,
2019,46(09):243-249.
[4] Lu Ping, Dong Husheng. Face image generation based on deep
convolution generation network [J]. Modern computer, 2019 (21):56-
58+64.
[5] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David
Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio, 2014,
Generative Adversarial Networks.
[6] Chen Lifang, Lu Guojun. MNIST handwritten digital image generation
method based on VAE and Gan fusion network [J]. Journal of Langfang
Normal University (NATURAL SCIENCE EDITION), 2019, 19 (02):
25-29.
[7] zhangjialiang, he Zhipeng, Wang Yuanyuan, Zeng Bing, Shen Yi, Jia
Yu. An improved method of fast image generation based on limited data
set [J]. Communication technology, 2019, 52 (05): 1119-1125.
[8] Zhang Guanghua, Wang Fubao, Duan Weijun. Astronomical image
generation based on GaN [J]. Journal of Northwest University of
technology, 2019,37(02):315-322.
[9] Alec Radford, Luke Metz, Soumith Chintala, 2015, Unsupervised
Representation Learning with Deep Convolutional Generative
Adversarial Networks.

50

Authorized licensed use limited to: University of Canberra. Downloaded on June 06,2020 at 16:15:18 UTC from IEEE Xplore. Restrictions apply.