Overview of Image Inpainting and Forensic Technology

Table 1

Comparison of deep CNN-based image inpainting methods.

Category	Model	Structure	Advantage/disadvantage	Perf.
CNN	Traditional CNN	Convolution layer; subsampling layer; full link layer	Efficient processing; automatic feature selection; good classification effect.
Deep CNN	Context encoder	Encoder and decoder structure combined with convolutional neural network	Unmarked image, irregular shape of missing area; can obtain the semantic information of the image; good repair effect, high processing speed.	17.58 dB (PSNR)
	PEN-NET	U-net structure with deep generative models	Visual and semantic coherence for image inpainting; fast convergence in training	9.94 (L1 loss)
	Partial convolution	Applying the partial convolution to replace all convolutions in the U-Net structure.	Stable performance; suitable for image inpainting with irregular shape.	19.04 dB (PSNR)
	Gated convolution	Using gated convolution to optimize partial convolution	Improved the flexibility; with mask and guiding input, the repair effect can be improved.	1.6% (mean L1 error)
	High-resolution	Content network and texture network	Consume a lot of computing resources and take a long time; only the patch in the picture is used instead of the data in the whole dataset.	18.00 dB (PSNR)
	Shift-net	U-net-based architecture	Image center restoration; continuity between pixels is not considered.	26.51 dB (PSNR)
	CSA	U-net architecture	Fast speed; good repair effect	26.54 dB (PSNR)
	RN	Encoder and decoder structure	Good repair effect	28.16 dB (PSNR)