The main purpose of this dataset is providing natural images to
test resampling algorithms (i.e. interpolation, zooming, enlargment
and superresolution).
This dataset uniquely provides a base of 40 RBG 2400x2400 reference
images with a bit-depth of 16bpp and high dynamic range (HDR).
The reference images were processed to obtain different
sub-resolutions and variation with 8bpp and grayscale, for a total
of more 220K images.
The test images were chosen to be square for simplicity and symmetry. The value 2400 was chosen because it is approximately the vertical resolution of a common 10 Megapixel SLR Camera, it is the vertical resolution of standard WQUXGA displays and can be divided by 2, 3, 4, 5, 6, 8, 10, 12, 15, 16, 20 and 24 (reduction factors).
Since currently there are no 16bpp capable photographic cameras in the consumer market, to acquire 16bpp natural images an approach involving multiple exposures was used.
This archive contain 40 16bpp RGB "uncropped" images with a resolution of 2448x2448 pixels. A border of 24 pixels was further cropped from these images to get the reference resolution of 2400x2400 pixels, so the top left pixel in the reference image correspond to the pixel in position (25,25) of the uncropped image.
For each reference image, additional resized images were generated. The intensity value of each pixel in the resized image was obtained by averaging the values of RxR sub-arrays (where R is the reducing factor). The reduction factors were chosen to perfectly divide the 2400 width and height values of the reference images:
| Reduction factor |
Resolution [Width x Height] |
Variations (shifts) |
|---|---|---|
| 1 | 2400x2400 | 1 |
| 2 | 1200x1200 | 9 |
| 3 | 800x800 | 25 |
| 4 | 600x600 | 49 |
| 5 | 480x480 | 81 |
| 6 | 400x400 | 121 |
| 8 | 300x300 | 225 |
| 10 | 240x240 | 361 |
| 12 | 200x200 | 529 |
| 15 | 160x160 | 841 |
| 16 | 150x150 | 961 |
| 20 | 120x120 | 1521 |
| 24 | 100x100 | 2209 |
The resized images were always generated starting from the
uncropped image but always cosidering a sub-sampling mask of
2400x2400 pixels at the time.
For each resizing resolution, all the possible variations obtained
by moving (shifting) the sub-sampling mask were saved.
The image names contain a shifting indicator, a string composed by
two groups of one letter and two numbers. The letters indicates the
direction of the shift (C: center, T: top, B: bottom, L: left, R:
right) while the numbers indicate the number of pixels. The images
aligned with the reference image, or the pixel in position (25,25)
of the uncropped image, are identified by the "C00C00" shifting
indicator.