2-D linear interpolation with bad pixel values

Frank Valdes wrote on Dec 17, 1998

Q:  I am interested in performing a 2-D linear interpolation on an image.  
The complication is that I want the interpolation to ignore pixels with values, say,
of zero.   I've looked at programs like magnify, but have not found the option to ignore
zero-valued pixels.  Do you have any suggestions?


A:  What you would like is desirable but the current versions of the
interpolation tasks do not provide the ability to ignore flagged
pixels.  I suggest you do the following.

1.  Make a mask (an image with .pl extension) of the pixels you want
to ignore.  If you already have them set to zero then use the command

cl> imexpr "a == 0 ? 1 : 0" mask.pl image

This makes mask.pl based on image where any value of 0 has a mask value
of 1 (which means a flagged or bad pixel) and zero otherwise.  If you
have real data you may need to use a looser test than a==0 such as
abs(a)<0.0001.

2.  Use FIXPIX to replace the flagged pixels by interpolation from
neighboring pixels.  You need to do this to avoid having the interpolator
do funny things around the zero valued pixels.

3. Do the interpolation.  If the interpolation changes the coordinates,
such as scale change or shifts, and you want to do step 4 below to
replace the original flagged pixels by zero again, do the following.
Set the mask values to a larger value (in step 1 use a value of 1000
instead of 1 or just use IMARITH to multiply).  Interpolate the mask
in exactly the same way as the image. [For large images there can be
a bug such that the output of the interpolation to a .pl output image
name is wrong.  In this case interpolate to a regular image as output
and then convert back to a mask (or simply leave it alone for step
4).]

4. You can then restore the flagged pixels to zero (if you want otherwise
the information is in the mask) with

cl> imexpr "b == 1 ? 0 : a" newimage image mask

where "image" is the interpolated image and mask is the mask image giving
the location of the flagged pixels.  This mask would either be the original
mask if the interpolation did not change the coordinates or the interpolated
mask if it did.

This is done with Mosaic data and works well (but the pixels are always
left at the interpolated values rather than the original values or a
flag value).  The interpolated masks are used to stack dithered images
so that the missing values are actually replaced with real data rather
than interpolation from surrounding data.

---

Last post on Dec 17, 1998