Deep CNN and Deep GAN in Computational Visual Perception-Driven Image Analysis

Table 7

Comparisons of images generated by D-GAN variants.

Type of D-GAN	Epoch 10	Training losses	Epoch 100	Training losses	Characteristics	Shortcomings
D-GAN					Generator-discriminator framework via a minimax game where samples are directly generated.	The model parameters oscillate, destabilize, and never converge.
DCGAN					It uses convolutional stride and transposed convolution for the downsampling and the upsampling.	Gradients disappear or explode.
CGAN					Conditional generation of images.	CGAN is not strictly unsupervised. Some labeling is required for it to work strictly.
LSGAN					It creates high-quality images compared to the GAN. It is more stable during training.	Additional computational cost.
WGAN					Stability of learning and overcomes the mode collapse problem.	The difficulty in the WGAN is to enforce the Lipschitz constraint.