Encyclopedia of Graphics File Formats Companion/GFF_CD.ISO/software/unix/saoimage/saodoc.tar/scale.txt

Text File | 1989-11-29 | 10KB | 192 lines

SAOimage scaling


scaling from image file values to the range of the display color map

the color map

The workstation display only associates colors with a small, finite set
of values (typically integer values between 0 and 255).  The values used
by the workstations are associated to the colors in its palette through
a color map (e.g. 0=black, 1=dark blue, ...).  Some of the color map
entries are reserved for such things as the colors of the screen or
window background, the mouse cursor, and terminal window text, leaving
fewer color map values for an application like SAOimage to use.  See the
color chapter for a more detailed discussion.

the image data

Image data may have an arbitrary range of data values.  Data values may
be very large or be very small or negative values, and they may have
fractional components.  Image scaling involves binning groups of image
data values for each display color map value.

basic scaling

Take the case of an image with a range of values from 0 to +1199 (a
range of 1200).  Imagine that SAOimage has only 200 color map values
to work with.  It might assign one color map slot (or bin) to each
range of six in the image.  Thus values 0 to 5 would be assigned the
lowest color map slot, say 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, but not so important to see a difference between 1001 and 1002.
Perhaps it would be sufficient to see differences between 1001 and 1100,
but 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.  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.  The square root function would be similar
but follow a different curve.

non-linear distributions

Suppose that weighting toward one end or the other wasn'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 80%
of the color cells would be used for non-existent values between 200 and
1000.  Worse, as is common with CCD images, perhaps there are bad pixels
with values like -1200, which have nothing to do with the data.

 equalization

It is appropriate to speak in terms of a histogram of the data values,
plotting ranges of data values by the frequency (number of pixels) of
their occurrence in the image.  Suppose that the ranges in the histogram
were selected such that each range was associated with one color map
value.  In a poor scaling distribution, many of the ranges have few or
no actual occurrences in the image.  Histogram equalization is a process
whereby each range is adjusted with respect to the others, until each
range has about the same number of pixel occurrences in the image, thus
maximizing the information in the display.  This usually produces a
dramatic improvement in the amount of visible detail in the display.

The drawback of histogram equalization is that the ranges can vary
greatly in size, thus 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

Where most of the data falls within a single contiguous range, with bad
pixels or infrequent events in the extremes to either end, scaling to the
color map can be restricted 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 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.

stretching the color map range

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 base range, 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 -32768 and 32767 (a range of 65,536).
Each scaling function produces a look-up table (or map) 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 (-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 look-up table mappings, depending on the contents
of the display.  Panning and zooming can be used in this way to find a
better scaling.  To allow for extreme data that just appear at the edges
(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 scaling can be restricted
by using the -min and -max arguments on the command line.  When given,
the low end of the range will not extend below the -min value, and/or
the upper end will not extend above the -max value.  The limit value
should be given in terms of original file values (not the internal
short-integer values, should they differ).  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
scaling map to be calculated.  The display is drawn with the existing
scaling map.  If the new display has values outside the range when the
map was made, these values are usually clipped  in terms of a histogram of the data values,
plotting ranges of data values by the frequency (number of pixels) of
their occurrence in the image.  Suppose that the ranges in the histogram
were selected such that each range was associated with one color map
value.  In a poor scaling distribution, many of the ranges have few or
no actual occurrences in the image.  Histogram equalization is a process
whereby each range is adjusted with respect to the others, until each
range has about the same number of pixel occurrences in the image, thus
maximizing the information in the display.  This usually produces a
dramatic improvement in the amount of visible detail in the display.

The drawback of histogram equalization is that the ranges can vary
greatly in size, thus 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

Where most of the data falls within a single contiguous range, with bad
pixels or infrequent events in the extremes to either end, scaling to the
color map can be restricted 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 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.

stretching the color map range

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 base range, 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 -32768 and 32767 (a range of 65,536).
Each scaling function produces a look-up table (or map) 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 (-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 t