Calibration Basics
THE
DIGITAL CORNER
I
had hoped to present a reasonable summary of the calibration process this month,
but find it to be too big a project to handle all at once, especially with my
current work schedule. So let's
begin with just an overview of the basics.
Calibration
in the context of digital photography generally means setting things up so that
the final output (in whatever form that might be) will look the way you intend.
A parallel, if not similar, process is or should be used in conventional
photography as well, and this has been described in one or more club meetings.
This process makes it possible to switch from one type of film to another
without having to make extensive tests each time.
Likewise, when you take film to a commercial processor for printing, they
(hopefully) use calibrated processes to produce prints that are very close to
the correct color and value scale. One
of the biggest problems in getting good digital images is that most people don't
understand the calibration process for either domain, and are further confused
by the differences and limitations of digital imaging.
One
can think of the calibration problem as having a "color" part and a
"value" part. Value is
easier to understand, but is more commonly the problem. If we deal only in black & white images (actually "greyscale"
images in the digital world), we need be concerned only with value.
Our original image has a certain density range (which expresses the ratio
of lightest to darkest regions) and a "transfer function", which
describes (via a graphed curve) how the density relates to linear values of
light. Ideally, this transfer
function would be a straight line and the ratio of lightest to darkest regions
would be enormous, so that every exposure would produce an image with detail
information in every part of the image. This
is not the case, of course. Film
typically provides a fairly straight-line transfer function over a good range of
light values, and gradually becomes "flat" at either end of the scale.
We already live with this limitation of film and know how to deal with
it.
In
the digital world "values" are expressed by numbers, and within a
given digital system the range of these numbers is limited to a specific set,
usually 0 to 255. This range
corresponds to the range of numbers that can be expressed with 8 bits (1 byte),
which is a very fundamental data unit. It
can also be argued that this range is appropriate because humans can (generally)
distinguish less than 256 separate values.
But this assumes that you have selected a suitable set of 256 values,
having an optimal range and transfer function.
When an image is scanned the exposure and transfer function of the
scanner must be matched to the density and transfer function of the film in
order to capture an optimal image. To
understand transfer function better we can describe it with three more common
terms: Brightness, Contrast, and
"Gamma". Gamma is a new term to some people, and is used somewhat
differently in different technology worlds.
What I mean by "gamma" is an exponential curve from black to
white (a "bow" that can be either above or below a straight line from
black to white). Contrast is the
slope (steepness) of the curve at the center.
The more vertical the line is, the more contrast.
Brightness (although this is a poor term for it) is the "base"
level - the level of the darkest part of the image.
Refer to the graphs and sample images on the next page.
Gamma
is particularly troublesome because people are not familiar with the concept and
mistake it for contrast and/or brightness.
Non-linear gamma is often an inherent limitation of a given imaging
technology. Various digital imaging
devices have different gammas. A
color monitor typically has a gamma of 1.8, which is pretty seriously non-linear
(linear would be a gamma of 1). One
of the first steps in calibrating a digital imaging system is to compensate for
the gamma of the monitor. This can
be done reasonably well by comparing a mid-range value with the average value of
a tight pattern of blacks and whites and adjusting a transfer function to make
them equal.
That
makes the display more or less linear, but do we want the image data itself to
be linear? In an ideal world, yes.
But in the real world we have to consider the gamma of the ultimate
output device. We could keep image
data in a linear form and then adjust it for a given output device just before
sending to that device. This may be necessary if the image will go to many devices or
if we don't know what device it will go to, but remember that we have only 256
levels to work with: Each time we
adjust the transfer function some levels are lost, which results in a loss of
detail. The more severe a gamma
correction, the more levels are lost. So
ideal output is achieved when the image is scanned and maintained in the gamma
"domain" of the intended output device.
This gamma is then compensated for in the computer display system (in
addition to compensation for the monitor itself), so that we still see a linear
output more-or-less representing what the final output will be like.
If the output device has a severe gamma (and depending on how it relates
to the monitor gamma), the display may show some contouring or loss of detail,
but this is usually not a problem and is better, in any case, than having such
problems in the final output.
After
the value scale is properly adjusted we can consider color.
Except for the real world, color images are almost always expressed using
a "tri-stimulus" model, meaning that the color at a given point is
represented by the intensity of three specific frequencies (hues) of light.
A color monitor uses the Red/Green/Blue set of primary colors, while
printing processes usually use the Cyan/Magenta/Yellow primaries.
Essentially, three colors are chosen to cover as broad a range as
possible. If the three hues are
marked on a chromaticity chart the triangle defined by these points contains all
the colors that can be created by various combinations of the three primaries.
It's important to note here that it makes a difference what the precise
hue of the primaries is. For example, if I scan a color photograph the scanner
separates the red, green, and blue using its built-in filters. When I display the image on a color monitor the colors are
roughly the same, but each channel of image data stimulates color phosphors on
the face of the monitor and the hue of the light emitted depends solely on the
properties of that phosphor. If
only one phosphor color doesn't match the corresponding filter the entire color
space is skewed. If the hue
difference is small most color points will still be within the triangle and thus
representable (particularly since "real world" images tend to have
very few highly saturated colors). But
mapping the input to the output requires more than adjusting the relative levels
of the primaries. Such a skew (or
rotation) of the color space can only be corrected by mathematically
"rotating" the color data. This
is done via a matrix multiplication: An input set [R,G,B] is multiplied by a calculated 3 by 3
matrix to produce the output set [R',G',B'].
The
same mathematical rotation (with more extreme coefficients) can be used to
transform an RGB image into a CMY image. Obviously,
the overlap between these two triangles is not good at all, with a great many of
the more saturated colors becoming impossible to reproduce.
As with the value scale, we must know the color space of the output
device to reasonably preview what the image will look like.
Fortunately, CMY output devices tend to have smaller color spaces than
RGB, so relatively few colors in the final output cannot be displayed on a color
monitor. However, a good deal of
the RGB color space cannot be reproduced in CMY, so it is important to preview
the converted image. You might
never be able to get the color you want, but at least you won't be surprised
when you see the final output.
Next
month we'll delve into some specifics of calibrating your system.
Copyright (C) 2004 Greg Marshall
No portion of this web site may be copied or reproduced in any manner without express written consent.