Overview of Image Inpainting and Forensic Technology

Table 2

Comparison of different types of structural change-based GAN.

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)
Conditional GAN	CGAN	Adding constraints to the input layer of GAN to guide generation	High requirements for the dataset; tagged dataset; low quality image.	—
	Info-GAN	Use generator and discriminator and add hidden code in the generator	Extending the theory and increasing the interpretability of GAN model; large amount of computation; and poor diversity.	—
Deep convolutional GAN	DCGAN	Symmetrical discriminator and the generator; using fractional-strided convolution instead of upsampling.	Stability training; No artifacts; clearer edges; data distribution affect the effect; unsuitable for complex scenarios.	85.95 (accuracy)
	GL	A inpainting network and two discriminators network	The result is locally and globally consistent; can complete any scene; and unsuitable for heavily structured objects.	96.5% (accuracy)
Combined GAN	DGM	Global and local discriminators; two stages feedforward generation network.	Expanding the receptive field and improving the stability of network training; less parameters.	18.91 (PSNR)
	StackGAN	Superimpose two CGANs	High-resolution images; stable training.	8.45 (IS)
	Cycle-GAN	Circular mechanism to achieve two domain conversions and constraints	Low data requirements; no one-to-one paired images; low resolution is not high.	0.58 (accuracy)
	PICnet	Parallel generating paths and reconstructing paths; variation encoder structure of the generator.	The reliability of complementary content; high-quality images；suitable for the images with arbitrary of incomplete parts.	20.10 (PSNR)
Pattern-based GAN	PGGAN	Gradually grow generators and discriminators.	Reducing the training time; generating better high-resolution images.	0.2838 (MS-SSIM)
	Style-GAN	Including mapping network and synthesis network; style-based structure of generator.	Can control the visual features from coarse features to fine details.	4.40 (FID)
	Style-GAN2	Replace the gradient network structure of style GAN with fixed network training.	Focus on repairing artifacts and further improving the quality of generated images	2.32 (FID)
Self-attention GAN	SAGAN	Both generative network and discriminant network adopt attention mechanism.	Getting good balance between improving the receptive field and reducing the amount of parameters.	18.65 (FID)
	BigGAN	More channels in convolutional layer and use truncation and orthogonal regularization	Model training is stable and can generate ultra-clear images; large amount of parameters; difficult to train	7.4 (FID)