Generative Adversarial Network Technologies and Applications in Computer Vision

Table 4

Comparisons of GAN models on the structure.

GAN models	Improvements	Shortages	Applications
MAD-GAN [49]	The model uses many generators and a discriminator to generate samples	Hard to convergence and lack of diversity	The model is applied to multivariate time series anomaly detection
InfoGAN [41]	The model's input is composed of c and z’, through adding a classifier to predict code c that generates x	Complex and with a large number of params	The model is unsupervised and learns interpretable and disentangled representations on challenging datasets
ACGAN [4]	The model combines the advantages of CGAN and SGAN to generate samples	Semisupervised and the model is hard to converge in the small amount of data	Can generate high-quality samples and have diversity
StackGAN [50]	The model through two-stage training generating more realistic samples	Complicated and need more training time	The model can be applied according to text to generate images