Generative Adversarial Network Technologies and Applications in Computer Vision

Table 5

Comparisons of GAN models in high-quality examples generation.

GAN models	Improvements	Shortages	Applications
DCGAN [63]	The methods fraction-strided convolution, batchnorm, and ReLU make the model more stable and easy to converge	Model collapses and needs to adjust parameters in different conditions	Highest usage models in most scenarios
LAPGAN [64]	Laplacian and Gaussian pyramids in the up and down samples which make the model easy to approach and learn residuals	Supervision model	High-resolution images generation
SAGAN [65]	Using self-attention mechanism and two-timescale update rule, the model can generate realistic images	The attention mechanism is limited	Large-scale classification of conditional image generation tasks
VSRResFeat GAN [61]	Using GAN loss and Charbonnier distance in feature and pixel space	The noise in the estimated frames is redundant	Video super-resolution