Synopsis
Workstation displays pose a common problem for imaging applications:
how to display an image with a large range of values on a screen
with a small range of colors. Since there aren't enough colors to use
one for each image value, groups of different image values must all be
assigned the same color. The process of grouping image values into a
smaller set of values is called scaling or binning.
Use the Scale submenu to scale the image to the range of display
colors. Any of a variety of linear and non-linear scaling algorithms
can be applied simply by clicking on the appropriate submenu button.
When scaling is performed, the range of image values currently visible
within the area of the display window will be assumed. Limits on the
range of image values and other scaling parameters may be input using
command line switches. For long integer and floating point data,
SAOimage uses two stages of scaling, as described below.
Range of Image Data vs. Size of Color Map
Image data may have a nearly infinite range of data values. Data
values may be very large or very small, they may be positive or
negative, and they may be real or integer.
A typical workstation can only display 256 colors on its screen at any
one time. Some of the 256 colors are needed for coloring such things
as the mouse cursor, text characters, background areas, and window
borders. On a 256 color workstation, SAOimage typically uses
200 colors to display images.
For purposes of display, each pixel in the display image is assigned
one of 200 integer values in the range of 0 to 255. Each of these
256 possible values is the index of an entry (or slot) in the
display's color map table. The color map table associates specific
colors with the 256 display values (e.g. 0=black, 1=dark blue, etc.).
(See the COLOR section for a detailed discussion of color map
manipulation.)
Basic Scaling Theory
Take the case of an image with a range of values from 0 to +1199 (a
range of 1200), and an image display with 200 colors. SAOimage
might assign one color map entry (or slot) to each range of six in the
image. Thus values 0 to 5 would be assigned the lowest color map
slot, 0. 6 to 11 would be assigned to 1, and so on up to the values
between 1195 and 1199 going to slot 199 in the color map. Then, if
slot 0 was black, all pixels in the image with data values between 0
and 5 would appear black in the display. If color map slot 131 was
assigned the color pink, then image pixels whose data value was
between 787 and 792 would appear pink in the display.
Scaling Distributions
The above example is an example of linear scaling. But suppose that it
was important to be able to see the differences among levels 2, 3, 4,
and 5, and not as important to see a difference between 1001 and 1002.
Perhaps it would be sufficient to see differences between 1001 and 1100,
while 1001, 1002, 1003, etc. could all have the same color. Then it would
be better to use a scaling that used more color map slots for the lower
image values than the higher image values. Log and square root scaling
do that.
To understand how that works, imagine a graph with image values along
the side and color map values along the bottom. A plot of linear
scaling would be a straight line with a constant slope. In other
words, each n consecutive image value units would map to m
color map levels. A log line would be curved, having less slope at
the bottom (more color map values per image value range) and get
steeper toward the top (fewer color map values per image value range).
The square root function would be similar but follow a different
curve.
Non-linear Distributions
Sometimes weighting toward one end or the other isn't good enough.
Perhaps the data values fall in clusters with big gaps in between.
Imagine an image with most of its data clustered between 0 and 200,
but two or three pixels with a value of 1000. In the linear scaling
80200 and 1000. Worse, as is common with CCD images, there may be bad
pixels with values like -1200, which have nothing to do with image
source values.
Histogram Equalization
A histogram is a plot with the range of possible values (divided into
discrete bins) on one axis and the frequency of the actual occurrence
of each value, or value within the range of each bin, on the other
axis
.
The image histogram could be divided into as many bins as colors in
the display, with each bin representing the range of image values
associated with a single color in the display. The frequency value
of each bin would be the number of pixels in the display that used
that color. In a poor scaling distribution, many of the bins (ranges
of image values) have few or no actual occurrences in the image.
Histogram equalization is a process of adjusting the ranges of the
bins such that each bin has about the same number of image pixels.
If the image has many pixels with values between 100 and 200, then
histogram equalization would allocate many small bins to that range,
(e.g. 100-110, 111-120, etc.). If the image has relatively few
pixels with values between 300 and 400, histogram equalization would
allocate few large bins to that range (e.g. 300-350, 351-400). The
object of histogram equalization is to maximize the information in
the display by optimizing the usage of the available colors. It
usually produces a dramatic improvement in the amount of visible
detail in the displayed image.
The drawback of histogram equalization is that the ranges can vary
greatly in size. A difference between two adjacent color map values
may represent a big or a small difference in data values, with much
irregularity from one step to the next.
Windowing the Scaling
Often, most of the data falls within a single contiguous range, with bad
pixels or infrequent events in the extremes to either end. It may be
useful to restrict the scaling algorithm to a specific range, with
all values below that range having the same color as the minimum and
all values above that range having the same color as the maximum.
This can either be accomplished by specifying the scaleing limits
directly, or by looking for a section of the image which is free from
extreme values, and basing the scaling on the values in that section
only.
Wrapping the Color Map
Another trick to show more detail with a limited color map is to reuse
it. By wrapping the color map on itself, it could represent the range
of values from 0 to 99, and then start over, 100 having the same color
as 0 and 199 having the same color as 99. This works for smoothly
varying data and even has a sort of contoured look to it. Where the
data fluctuations are greater than the range covered by one wrap of
the color map, the result is a lot of meaningless noise.
Realization and Efficiency within SAOimage
In order to realize reasonable execution times for rescaling the image
display, SAOimage does not apply its scaling functions to real
data. Where the image data was real, double, or long, the data is
linearly rescaled to integer values between -32767 and 32767 (a range
of 65,535). Each scaling function produces a 65,535 entry lookup
table for this range of values which is used to draw or redraw the
display. If the data has such great extremes that this range linearly
applied to the data's range will be insufficient, windowing limits for
the initial linear scaling can be given on the command line using
-rmin, -rmax.
The scaling function is further windowed by being applied only to the
range of image values actually being displayed at the time the
function is invoked. Pixels not appearing in the display window are
not considered in the scaling algorithm. Thus scaling after panning
and zooming will produce different scale lookup table mappings,
depending on the contents of the display. Panning and zooming can be
used in this way to find a better scaling.
To compensate for extreme data that may appear at the edges of an image
(common in CCD images) data within 2 pixels of the edge of the display
are not considered in the assessment of range. In verbose mode,
the scaling routine reports the range of values which it finds in the
display.
One or both limits of the value range for menu-commanded scaling can
be restricted by using the -min and -max arguments on the
command line. When -min is given, the low end of the range will
not extend below the -min value. When -max is given, the
upper end will not extend above the -max value. Limit value are
given in terms of original file values. The user need not know about
the internally used short integer values. The -min and/or
-max value can be cleared by giving the -min or -max
switch with no argument.
Panning or zooming after a scale map has been made does not cause a new
scale map to be calculated. The display is drawn with the existing
scaling. If the new display has values outside the range when the map
was made, these values are usually clipped (mapped to the color map
minimum or maximum). Wrapped scaling is an exception, in that the
color map wrapping is applied all the way to the maximum possible
value.
Linear Scaling
Linear scaling is fairly simple. The range of image values is divided by
the number of color map values to determine the range of data values for
each color map value. The fixed range is applied to mapping all image
values between the minimum and maximum image value. Values below the
minimum are all mapped to the lowest color map value. Values above
the maximum are all mapped to the highest color map value.
Wrapped Linear Scaling
Wrapped linear scaling, as described above, divides the range of data
values by the number of times the color table will roll over (the
default is 10) and the result is divided by the number of color map
values to determine the range of data values for each color map value.
Values below the minimum are all mapped to the lowest color map value.
The mapping continues to be rolled up to the highest map value
(32767). (see -wrap in the COMMAND LINE section)
Log Scaling
The distribution of data values to color map values is based on the
distribution of e**n from 0 to X, where n is a
parameter (the default is 10.0) and X is determined by n
and the two ranges (data and color map). Positive values of n
favor the lower data values, while negative values of n favor
the higher data values. Values below the minimum are all mapped
to the lowest color map value. Values above the maximum are all mapped to
the highest color map value. (see -log in the COMMAND LINE
section)
Square Root Scaling
The distribution of data values to color map values is based on the
distribution of X**(1/n) from 0 to 1, where n is a
parameter (the default is 2.0) and X varies from 0 to 1 in steps
determined by the number of color map values. Values of n
greater than 1 favor the lower data values, while values of n
less than 1 favor the higher data values (negative values and 0 are
not allowed). Values below the minimum are all mapped to the lowest
color map value. Values above the maximum are all mapped to the
highest color map value. (see -sqrt in the COMMAND LINE
section)
Histogram Equalization
The principle of histogram equalization is described above. The
histogram equalization algorithm in SAOimage differs from common
one-pass algorithms by accounting for disproportionately large
occurrences of one or a few image values. In an image with 1000
pixels, 500 of which have the value 32, a one-pass algorithm, given
100 colors, will try to get 10 pixels for each color. It will end up
with 49 unused colors, since one color covered 500 pixels.
SAOimage's algorithm detects the peak (or peaks) in the histogram
and allocates the remaining 99 colors among the 500 remaining pixels
(5 pixels per color).
IRAF
Images sent directly by IRAF are already scaled to the range
1-200. IRAF's own scaling defaults to a windowed linear scaling
where the window is determined by fitting a straight line to a small,
200 pixel, subsampled histogram. Further rescaling within
SAOimage is not very useful (see the IRAF section).