Page 34 - Read Online

Bao et al. Complex Eng Syst 2022;2:16 I http://dx.doi.org/10.20517/ces.2022.30 Page 3 of 10

nition accuracy. In addition, since the residual network [17–19] is able to overcome the gradient disappearance
problem in the deep neural network model, a deep neural network model containing intra-block residual
connections and inter-block residual connections was included in the proposed model during the pulsar can-
didate identification stage. Experiments proved that the module achieved optimal recognition accuracy in all
scenarios compared to the shallow deep neural network model.

In this paper, two basic methods utilized in the proposed model are first introduced with detailed figures
and illustrations. These basic methods are the generative adversarial network-based pulsar image generation
method and the residual network-based pulsar candidate identification method. Then, the HTRU-Medlat
dataset is used in the proposed model and the experimental results are obtained; these results indicate that the
proposed model achieves the best performance in the experiment without any other complicated data genera-
tionmethod,whichiswhyHTRU-Medlatwaschosenasthedatasetfortheexperiment. Intheproposedmodel,
a time-versus-phase plot and a frequency-versus-phase plot are used to implement the screening of pulsar can-
didates and describe their characteristics of pulsar candidates so that samples can be better evaluated. As a
result, positive samples can be identified more accurately, which is essential because positive samples represent
the essential information of pulsar candidates. After collecting the experimental results, several evaluation in-
dicators, including Precision, Recall and F1-score, are selected to assess our experimental results. Finally, the
results of the proposed models are compared with other existing models. The experimental results show that
ResNet exhibits a better ability to fit data and extract features on small datasets and large datasets containing
images generated by generative adversarial networks compared to CNN methods. In the comparison experi-
ments between the small dataset and the extended large dataset, the improved F1 values and accuracy metrics
of the CNN method indicate that the simulated sample-extended data generated by the generative adversarial
network can improve the model’s accuracy to some extent. Through experimental validation, better results
are obtained: the quality of the large dataset extended with simulated samples is improved, providing richer
[4]
recognition features, and the recognition accuracy is further improved .

2. METHODS
In this paper, a framework combining generative adversarial networks and residual networks is proposed for
pulsar candidate identification. First, the generative adversarial networks module is used to tackle the imbal-
ance problem in the dataset, and it is able to generate a series of pulsar candidate images that approximate
positive samples to expand the existing pulsar dataset. Then, based on the idea of residual connectivity, this
paper designs a deep neural network for pulsar candidate identification utilizing intra- and inter-block residual
connectivity, which can effectively improve the recognition accuracy.

2.1. Generative adversarial network-based pulsar image generation method
In this paper, the Wasserstein distance [19] is used to replace the Jensen-Shannon (JS) divergence of the tra-
ditional GAN [20,21] because it can more explicitly measure the difference between two different distributions.
Wasserstein GAN (WGAN) based on the Wasserstein distance can overcome problems such as the gradient
vanishing and mode collapse experienced by the traditional GAN training, and it can generate more stable
pulsar images.

The architecture of the generator is illustrated in Figure 1. The generator produces 1 × 48 × 48 grey images by
accepting 1 × 1024-dimensional Gaussian noise as inputs. First, the Gaussian noise is transformed into a 1 ×
18432 tensor with a 1024 × 18432 fully connected layer. Then, it is projected and reshaped into a 128 × 12 × 12
tensor with 128 channels. In the next two convolutions, kernels of size 4 × 4 are adopted and the numbers of
channels are 64 and 1. The LeakyReLU activation function with a slope of 0.1 is used for all of the convolutions
except the last one, and the sigmoid activation function is used in the last convolutional layer to ensure that
the pixel values of the final output are in the range [0, 1] to generate appropriate grey images.