Image Resampling

Written by Jonathan Sachs

Copyright © 2001 Digital Light & Color

What is Resampling?
Digital images consist of a rectangular grid of evenly spaced pixels. Each pixel can
be thought of as a measurement or sample of the light from a subject. Commonly,
the original samples are obtained using a scanner or digital camera by averaging the
amount of red, green and blue light that falls on the sensitive area of each of its
CCD sensing elements.

Resampling is the mathematical technique used to create a new version of the

image with a different width and/or height in pixels. Increasing the size of an image
is called upsampling; reducing its size is called downsampling.

When images are upsampled, the number of pixels increases, but, with reference to
the original subject, new image detail cannot be created that was not already present
in the original image. As a result, images normally become softer the more they are
enlarged since the amount of information per pixel goes down.

When images are downsampled, information in the original image has to be dis-
carded to make the image smaller. Thus if you downsample and then upsample an
image, you will not get all the original image detail back. Downsampling a soft
image can make it appear sharper even though it contains less information than the
original.

Image Resampling 1
 Image Resampling

Resampling in Picture Window is performed by the Resize transformation. The

Resize dialog box displays the input image width, height and dpi setting and lets
you specify the output image width, height, dpi setting and resampling method.
Note that changing the dpi setting (resolution) does not affect the image data in any
way—it is simply a suggested scale factor that is associated with the image to expe-
dite printing or document the size of the original subject.

Why is Resampling Important?

A few situations where image resampling comes into play are:
• Reducing the size of an image from a scanner or digital camera for emailing or
display on the web.
This step is normally followed by saving the file in JPEG format and uploading
it to a web site or attaching it to an email message.
• Increasing the size of an image before or during the printing process.
Printing an image almost always requires resampling it as it will usually be too
small or too large to make a print of the desired size. This resampling is handled
automatically by your Windows printer driver. Most drivers do a good job of
resizing images, but you can resample the image yourself prior to printing if you
suspect the driver is not doing a good job.
• Creating Windows wallpaper
Wallpaper is a term used to describe an image that is used as a background for
the Windows desktop. You can create wallpaper from your own images by crop-
ping them to proportions of 4:3, resizing them to your current screen size (e.g.
1024x768) and them saving them as a (Version 1) BMP file in your Windows
folder. The file will then be available in the image list displayed in the Windows
Display control panel under Background.

2 Image Resampling
How Resampling Works

How Resampling Works

It turns out that resampling a 2-dimensional image can be broken down into two
one-dimensional resampling passes. In one pass, horizontal resampling is per-
formed producing an image with a different width but the same height. In the next
pass, this intermediate image is resampled vertically, changing it height while leav-
ing the width the same. This is computationally much more efficient than trying to
combine the work into a single pass.

Upsampling involves interpolating between the existing pixels to obtain an estimate

of their values at the new pixel locations.

Downsampling involves computing a weighted average of the original pixels that

overlap each new pixel.

Color images are treated like three black and white images which are separated
from the original image, individually resampled, and finally recombined to create
the final image.

Resampling Methods
Many different resampling schemes are possible. Most techniques work by comput-
ing new pixels as a weighted average of the surrounding pixels. The weights depend
on the distance between the new pixel location and the neighboring pixels. The sim-
plest methods consider only the immediate neighbors; more advanced methods
examine more of the surround pixels to attempt to produce a more accurate result.
Picture Window’s Resize transformation (starting with Version 3.0d) supports the
resampling methods listed below. Each one is illustrated with a small sample image
which is magnified by 1000% (i.e. enlarged by a factor of 10). This huge amount of
magnification is much greater than you would use under normal circumstances—it
is used here to make some of the relatively subtle differences between the different
methods more apparent.

Image Resampling 3
 Image Resampling

Nearest Neighbor

The nearest neighbor resampling method is very simple; it computes new pixels as
the value of the nearest pixel in the original image. This produces a blocky result
when upsampling and a coarse, grainy effect when downsampling.

4 Image Resampling
Resampling Methods

Bilinear

Bilinear resampling computes new pixels using linear interpolation. When upsam-
pling, this method operates on the 2 by 2 cell of pixels surrounding each new pixel
location. The resulting images are much smoother than those produced by the near-
est neighbor method but are not as sharp or as smooth as image computed using
bicubic resampling (see below).

Image Resampling 5
 Image Resampling

Bicubic
Bicubic resampling computes new pixels using a cubic splines. When upsampling,
this method operates on the 4 by 4 cell of pixels surrounding each new pixel loca-
tion. This is the recommended resampling method for most images as it represents
a good trade-off between accuracy and speed

0% Sharpening

16.7% Sharpening

25% Sharpening

This method has an extra parameter you can use to sharpen the image during the
resampling process. No sharpening gives the smoothest results; excessive sharpen-
ing can produce halos and more jagged diagonals. The default sharpening value of
16.7% is usually a good compromise.

6 Image Resampling
Resampling Methods

Lanczos 4X4, 6X6 and 8X8

Lanczos resampling computes new pixels using a Lanczos-windowed sinc kernal.

When upsampling, these methods operate on the 4 by 4, 6 by 6, or 8 by 8 cell of
pixels surrounding each new pixel location. The 4 X4 method is nearly identical to
the bicubic method with no sharpening; the 6X6 and 8X8 methods produce slightly
more accurate results than bicubic but take a little longer to compute.

4X4

6X6

8X8

Image Resampling 7
 Image Resampling

Resampling Kernals
Each of the resampling methods can be characterized by its resampling kernal. The
resampling kernal is simply a curve that determines the relative weights of adjacent
pixels based on their distance from the new pixel. The various kernals are repre-
sented by the following graph:

Resampling Kernals

Bicubic (0%) Bicubic (17%) Bicubic (33%)

Lanczos 4x4 Lanczos 6x6 Lanczos 8x8
Nearest Neighbor Bilinear

1.0
0.8
0.6
0.4
Weight

0.2
0.0
-0.2
-0.4
0 1 2 3 4
Pixels

8 Image Resampling
Genuine Fractals

Genuine Fractals
An alternative to the standard resampling techniques is a Photoshop plug-in called
Genuine Fractals from Altamira. This program does not resample images directly,
but it does let you save them using their proprietary fractal compression algorithms
and a special file format. When you reload the compressed file, you have the option
of resizing the image when it is opened and this in effect provides a fractal resam-
pling feature. Genuine Fractals has been hailed by some as a major breakthrough
for enlarging low resolution images from digital cameras so they can be printed at
large sizes without becoming fuzzy. Actual results seem to depend quite a bit on the
subject matter. The fractal images appear to be built up as areas of more of less
solid color separated by relatively sharp edges, a little like a watercolor painting.
The following examples compare fractal resampling to traditional resampling for
10X magnification

Genuine Fractals

Lanczos 8X8

Image Resampling 9
 Image Resampling

Here are some more samples of fractal vs. traditional resampling with a 10X mag-
nification:
Genuine Fractals

10 Image Resampling
Genuine Fractals

Lanczos 8X8

Image Resampling 11
 Image Resampling

Fractal Resampling

12 Image Resampling
Genuine Fractals

Lanczos 8X8

Image Resampling 13
 Image Resampling

Depending on the subject matter, it appears that fractal resampling is sometimes

fairly effective, sometimes produces results that are visibly different from tradi-
tional resampling but not necessarily better, and sometimes produces results that
are very similar. Whether or not you find the type of artifacts fractal resampling
introduces more or less pleasing is probably a matter of personal taste. Interest-
ingly, the first image (for which the results seem fairly good) is a tiny piece of a
Photo CD scan while the second two images are from digital cameras. The noise in
the digital camera images seems to reduce the effectiveness of the fractal resam-
pling scheme somewhat as there are fewer areas of nearly solid color and a lot of
false edges.

14 Image Resampling