Research into an Image Inpainting Algorithm via Multilevel Attention Progression Mechanism

Table 2

Quantitative comparisons on CelebA-HQ, Facade, and Places2.

Dataset	Mask rate	PSNR			SSIM			Mean L1 loss
		CA	MC	Ours	CA	MC	Ours	CA	MC	Ours
CelebA-HQ	10%–20%	26.16	29.61	31.33	0.9011	0.9330	0.9450	0.0380	0.0220	0.0182
	20%–30%	23.03	26.52	28.33	0.8351	0.8887	0.9082	0.0660	0.0378	0.0312
	30%–40%	21.62	24.92	26.93	0.7860	0.8552	0.8823	0.0872	0.0510	0.0402
	40%–50%	20.21	23.05	25.43	0.7255	0.8091	0.8492	0.1152	0.0692	0.0523
Facade	10%–20%	25.91	27.02	28.25	0.8975	0.9121	0.9265	0.0390	0.0320	0.0282
	20%–30%	25.31	24.48	25.32	0.8702	0.8575	0.8715	0.0645	0.0522	0.0472
	30%–40%	21.98	23.20	24.51	0.7802	0.8152	0.8415	0.0842	0.0682	0.0592
	40%–50%	20.88	21.91	23.31	0.7291	0.7702	0.8032	0.1060	0.0869	0.0742
Places2	10%–20%	22.48	27.35	27.67	0.8672	0.9102	0.9121	0.0592	0.0312	0.0292
	20%–30%	19.95	24.48	25.02	0.7862	0.8542	0.8570	0.0971	0.0512	0.0483
	30%–40%	18.56	22.70	23.40	0.7240	0.8002	0.8050	0.1312	0.0715	0.0660
	40%–50%	17.52	21.40	22.28	0.6588	0.7551	0.7652	0.1592	0.0892	0.0812