Epitomic Image Super-Resolution

Please download the MATLAB Executable (Win64 bit version) for Epitimic SR Here [Code]

Bird Image, input resolution 241 by 141, upscaled by 3 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Chip Image, input resolution 244 by 200, upscaled by 4 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Grass Image, input resolution 1280 by 720, upscaled by 3 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Ground Truth (GT)

Kid Image, input resolution 128 by 128, upscaled by 4 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Ground Truth (GT)

Leopard Image, input resolution 1280 by 720, upscaled by 3 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Ground Truth (GT)

Statue Image, input resolution 200 by 200, upscaled by 4 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Temple Image, input resolution 214 by 143, upscaled by 3 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Ground Truth (GT)

Texture Image, input resolution 166 by 166, upscaled by 3 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Ground Truth (GT)

Train Image, input resolution 321 by 214, upscaled by 3 times

Input

Bicubic Interpolation (Bicubic)

High Frequency Transfer (HFT)

Epitomic Super-Resolution (ESR)

Ground Truth (GT)

Subjective Evaluation

A pair of high resolution images produced by two different SR methods, for the same input low resolution image, are shown to each participant of this subjective survey. The participant selected which one is better than the other in terms of perceptual quality based on his/her subjective judgement. The high resolution image pairs are drawn from all the competitive methods randomly, and the images winning the pairwise comparison will be compared again in the next round, until the best one is selected. We use the Bradley-Terry model [1] to estimate the subjective scores for each SR method. 101 people participated the survey, and they performed such pairwise comparison over six images ( Kid, Chip, Statue, Leopard, Temple and Train).

Reference

[1] R. A. Bradley, and M. E. Terry, ``Rank analysis of incomplete block designs: I. The method of paired comparisons, Biometrika, pp. 324-345, 1952.