Overview of Image Inpainting and Forensic Technology

Table 3

Comparison of GAN model based on variant of loss function.

Category	Model	Structural features	Advantage/disadvantage	Perf.
Initial model	GAN	Generator and discriminator	High-definition images; unstable training; mode collapses; disappearing gradient.	225 (mean log likelihood)
GAN model based on loss function variation	LSGAN	Replacing the cross-entropy loss function with the least square loss function	Limit the unlimited modeling capabilities of GAN; unsolved the problem of gradient dispersion in the generator.	6.47 (IS)
	WGAN	Using Wasserstein distance instead of JS divergence	Stable training; theoretically solves the model collapse; gradients disappear and explode; slow convergence speed.	—
	WGAN-GP	Using gradient penalty mechanism instead of weight interception operation	Stable training; solves the problem of gradient disappearance and explosion; higher quality samples; slow convergence speed.	7.86 (IS)
	CASI	Fully convolutional network; introduced perceptual loss; joint loss function.	Reduces semantic errors; generates images that conform to the contextual content.	20.37 dB (PSNR)