A Generative Image Inpainting Model Based on Edge and Feature Self-Arrangement Constraints

Table 1

The network parameters of the edge generator.

Layer	Inputs (H × W × C)	Kernel size	Stride	Padding	Dilation	Activation function	Outputs (H × W × C)
ReflectionPad	256 × 256 × 3	—	—	3	—	—	262 × 262 × 3
Conv	262 × 262 × 3	7 × 7	1	0	1	ReLU	256 × 256 × 64
Conv	256 × 256 × 64	4 × 4	2	1	1	ReLU	128 × 128 × 128
Conv	128 × 128 × 128	4 × 4	2	1	1	ReLU	64 × 64 × 256
Residual Blocks × 8
ReflectionPad	64 × 64 × 256	—	—	2	—	—	68 × 68 × 256
Res-conv	68 × 68 × 256	3 × 3	1	0	2	ReLU	64 × 64 × 256
ReflectionPad	64 × 64 × 256	—	—	1	—	—	66 × 66 × 256
Res-conv	66 × 66 × 256	3 × 3	1	0	1	—	64 × 64 × 256
ConvTranspose	64 × 64 × 256	4 × 4	2	1	1	ReLU	128 × 128 × 128
ConvTranspose	128 × 128 × 128	4 × 4	2	1	1	ReLU	256 × 256 × 64
ReflectionPad	256 × 256 × 64	—	—	3	—	ReLU	262 × 262 × 64
Conv	262 × 262 × 64	7 × 7	1	0	1	—	256 × 256 × 1