Generative Adversarial Network Technologies and Applications in Computer Vision

Table 3

Comparisons of GAN models on the loss function.

GAN models	Improvements	Shortages	Applications
CGAN [8]	Through adding a conditional variable c to guide data generation and make the model faster to converge	The model is limited by data set and data set needs both tags and markeds	The model through semisupervised learning to generate a specified target
PAN [40]	The loss function was composed by the perceptual adversarial loss to train models	The model is a supervised model also needs the data set with both tags and markeds	The model can be applied to many image-to-image conversions
CWGAN [44]	The model is based on Wasserstein’s condition which has a lower cost than traditional GANs	Model collapses and has a lack of diversity	The model can be applied to short data set's training
FittingGAN [46]	The model is based on the CGAN loss function but adds an L1 regularization	The model accuracy is not very high and has a lack of diversity	Be better than the image-to-image task, it can generate images different from the input image guide