Fast Direct Super-Resolution by Simple Functions

Chih-Yuan Yang and Ming-Hsuan Yang

Electrical Engineering and Computer Science

University of California, Merced, USA

[paper] [supplementary material] [poster] [slides] [spotlight slide] [code] [bibtex] [FAQ]

Comparison

Code Release Version

09/01/2013 v1.0 First release
09/06/2013 v1.1 Update of a missing file F14c_Img2Grad_fast_suppressboundary
12/12/2013 v1.2 Release of re-organized test code, training images, a set of training code
05/03/2014 v1.3 Release of pre-trained priors for scaling factors 2, 3, 4, 5, 6, 8, and sigma values 0.4, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0.
06/30/2014 v1.4 size 24Gb. Release test images for scaling factors 2, 3, 4, 5, 6, 8, and sigma values 0.4, 0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8, 2.0.

Download sites: [Baidu Pan sharing folder] [Google Drive sharing folder]

Bibtex

@InProceedings{Yang13_ICCV_Fast,
author = {Chih-Yuan Yang and Ming-Hsuan Yang},
title = {Fast Direct Super-Resolution by Simple Functions},
booktitle = {Proceedings of IEEE International Conference on Computer Vision},
year = {2013},
}

Frequently Asked Questions

Q: On the second page of the paper, while computing the regression coefficients C, why its dimension is n*(m+1), rather than n*m?
A: The dimension is n*(m+1) rather than n*m because we extend the feature vectors by adding a constant value 1 at the end to compute an offset value in the regressed high-resolution features. Such a technique is used in linear regression or coordinate transformation. The value 1 is shown as the symbol 1 in Eq. 2.

Q: Why are four corners of the LR patch discarded?
A: The four corners are discarded in our design because we think they carry little useful information. They are far from the central predicted region, and highly affected by outside pixels where information is lost during patch cropping.

Q: As you select the center 12*12 pixels (rather than 16*16 or the whole patch) of the HR patch, does this selection have its theoretical basis or is it a conclusion of many experiments?
A: There is no theoretical basis for the selection of the 12*12 pixels. It is an empirical selection of a few issues. A low-resolution patch with an odd pixel number in width is easier to design because there is a central pixel. A Low-resolution patch in 3*3 or 5*5 is too small, thus we choose 7*7. Since boundary pixels are noisy, the proposed regression model is best to predict the central region, i.e. the central 3x3 in low-resolution. Since we want to validate the algorithm for a scaling factor of 4, the central 3x3 pixels in low-resolution mean 12*12 pixels in high-resolution. That is the reason the 12*12 pixels come.

Q: Where is the training image set containing 6152 files and 7.6 GB in size?
A: The compressed zip file is ftp://vision-U2.myftp.org/ICCV13/v1.2/Dataset/TrainingSet_zipped/AllFive.zip (single file in 7.6GB). If you are only interested in part of them, uncompressed files are in the another folder ftp://vision-U2.myftp.org/ICCV13/v1.2/Dataset/AllFive (6152 files).