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

Table 4

Quantitative comparison of algorithms.

Data set	Mask rate	Mean L1 loss				PSNR¶				SSIM¶
		GL	PConv	PEN-Net	Ours	GL	PConv	PEN-Net	Ours	GL	PConv	PEN-Net	Ours
Celeba	10% ∼ 20%	0.035	0.031	0.026	0.012	25.11	26.29	27.69	31.65	0.853	0.885	0.912	0.931
	20% ∼ 30%	0.062	0.051	0.041	0.036	23.96	24.21	26.44	30.21	0.811	0.817	0.826	0.839
	30% ∼ 40%	0.081	0.055	0.051	0.037	23.64	23.85	23.86	27.03	0.795	0.811	0.811	0.841
	40% ∼ 50%	0.101	0.089	0.072	0.065	22.21	21.47	22.37	26.41	0.713	0.786	0.797	0.823
Facade	10% ∼ 20%	0.043	0.033	0.031	0.019	23.08	23.59	24.19	32.36	0.801	0.857	0.841	0.943
	20% ∼ 30%	0.047	0.073	0.045	0.044	23.57	23.21	23.91	27.17	0.774	0.863	0.861	0.889
	30% ∼ 40%	0.088	0.091	0.073	0.055	21.15	22.19	22.57	25.33	0.766	0.712	0.854	0.871
	40% ∼ 50%	0.101	0.113	0.076	0.067	21.67	22.31	21.34	23.57	0.714	0.751	0.836	0.857
Places2	10% ∼ 20%	0.037	0.021	0.023	0.011	21.61	23.61	25.33	30.19	0.784	0.835	0.814	0.919
	20% ∼ 30%	0.057	0.054	0.040	0.031	21.17	22.71	24.74	27.47	0.776	0.797	0.804	0.814
	30% ∼ 40%	0.051	0.058	0.053	0.046	20.53	22.15	23.87	25.31	0.751	0.779	0.789	0.826
	40% ∼ 50%	0.071	0.083	0.075	0.066	20.27	21.94	21.61	22.92	0.737	0.764	0.778	0.793

The bold values are the experimental results in this paper.