Skip to content

Super resolution benchmark

WolframRhodium edited this page Mar 25, 2019 · 6 revisions

Super-resolution benchmark

Experimental settings

Testing Dataset: 100 private anime images, 1920 x 1080

SR Algorithms: Super-Resolution-Zoo (b575c32), znedi3

Evaluation methods: PSNR, WaDIQaM (FR/NR, dataset="tid", top="weighted")

A RGB high-resolution (HR) image is first downscaled by a factor of 2 using Catmull-Rom (resize.Bicubic(filter_param_a=0, filter_param_b=0.5)), then the low-resolution (LR) image is upscaled by a factor of 2 using a super-resolution algorithm (patch_size=image.size), resulting in a super-resolved (SR) image.

The chroma of the SR image is upscaled by catmull-rom for super-resolution algorithms that operate on luma channel (denote as Y).

The WaDIQaM_FR and PSNR scores are evaluated between the cropped (left=16, right=16, top=12, bottom=12) HR-SR image pair. PSNR is evaluated in luma color space.

The WaDIQaM_NR score is evaluated on the cropped (left=16, right=16, top=12, bottom=12) SR image.

Results (Rank by PSNR)

Algorithm PSNR, higher is better (std) WaDIQaM_FR, lower is better (std) WaDIQaM_NR, lower is better (std)
Origin 100.000 (0.000) 11.915 (1.253) 37.828 (11.788)
EDSR (baseline) 48.481 (3.230) 11.934 (1.875) 38.609 (11.782)
MDSR (baseline) 48.214 (3.241) 12.050 (2.078) 38.623 (11.771)
CARN 47.825 (3.139) 11.871 (2.185) 38.593 (11.824)
MR-ESPCN (Y) 47.717 (3.506) 12.392 (2.090) 38.217 (11.750)
MDSR (baseline, jpeg) 47.704 (3.201) 12.103 (1.866) 39.026 (11.759)
CARN-Mobile 47.025 (3.378) 12.231 (2.222) 38.706 (11.760)
waifu2x (upconv7, noise=-1) 45.466 (3.397) 14.745 (3.074) 40.570 (11.637)
IDN (Y) 45.250 (3.511) 14.832 (2.479) 40.242 (11.871)
waifu2x (upconv7, noise=0) 44.301 (3.170) 14.446 (3.723) 41.632 (11.104)
Catmull-Rom (invks) 44.110 (4.278) 16.543 (3.336) 40.357 (11.149)
nnedi3 (Y, nsize=0, nns=3, qual=2) 43.881 (4.005) 17.351 (3.366) 42.119 (11.136)
Lanczos4 43.253 (4.286) 18.649 (4.077) 42.227 (11.135)
Lanczos4 (noring) 43.013 (4.272) 18.924 (4.251) 42.485 (11.163)
Lanczos3 42.905 (4.315) 19.206 (4.353) 42.452 (10.946)
Lanczos3 (noring) 42.756 (4.295) 19.323 (4.427) 42.633 (11.037)
Spline36 42.569 (4.330) 19.843 (4.603) 42.917 (10.904)
Catmull-Rom 41.382 (4.344) 21.855 (5.270) 44.157 (10.887)

Please note that the results may be biased (model conversion may be wrong, the experiment setting is ideal, etc).