Table I.

Image Quality Assessment for synthetic experiments. We run ablation study to investigate the performance of models employing either Voronoi (2, 4) or Bicubic kernel (3, 5) for simulated noise or noise-free data. We compare Video ZSSR with Voronoi kernel (6) against SISR trained on synthetic data (7), state-of-the-art DCNN (8) [20], baseline interpolation reference method (2).

model	baseline	ZSSR					SISR	DCNN [20]
training	interpolation	noise-free frame	noisy frame	noisy video	simulation	pre-trained
kernel	LR (1)	Voronoi (2)	Cartesian (3)	Voronoi (4)	Cartesian (5)	Voronoi (6)	Voronoi (7)	Cartesian (8)
PSNR	27.99 ± 1.08	28.86 ± 1.08	28.27 ± 0.99	30.16 ± 1.22	28.23 ± 0.93	30.67 ± 1.27	30.99 ± 1.29	28.04 ± 1.08
SSIM	0.851 ± 0.020	0.880 ± 0.017	0.862 ± 0.014	0.878 ± 0.015	0.849 ± 0.021	0.890 ± 0.014	0.902 ± 0.012	0.852 ± 0.020
LIPIPS	0.781 ± 0.017	0.806 ± 0.015	0.785 ± 0.012	0.806 ± 0.014	0.777 ± 0.018	0.817 ± 0.012	0.816 ± 0.014	0.782 ± 0.017
GMSD	0.940 ± 0.006	0.954 ± 0.007	0.953 ± 0.004	0.955 ± 0.004	0.943 ± 0.006	0.960 ± 0.004	0.961 ± 0.004	0.940 ± 0.006
L1 loss	0.973 ± 0.003	0.975 ± 0.003	0.974 ± 0.003	0.980 ± 0.003	0.973 ± 0.003	0.981 ± 0.003	0.982 ± 0.003	0.973 ± 0.003
VGG	2.62 ± 0.19	2.39 ± 0.17	2.70 ± 0.14	2.37 ± 0.16	2.69 ± 0.20	2.24 ± 0.14	2.25 ± 0.16	2.60 ± 0.19