[gnuastro-devel] [task #14362] Automatic pixel value type on image manipulation

[Mohammad Akhlaghi]

Follow-up Comment #1, task #14362 (project gnuastro):

This is a great suggestion, thanks Lee.

Indeed the current behavior where the default ImageWarp is in the same type as
the input must be modified. For the time being, I am thinking of keeping the
double precision floating point as the default output and using a `--type'
option to allow the user to set the output type as they like (similar to what
is implemented in MakeProfiles, see MakeProfiles options
<https://www.gnu.org/software/gnuastro/manual/html_node/MakeProfiles-options.html>,
under the `--type' option). I will do this while correcting it as part of task
#14245 (I haven't got to ImageWarp yet!).

But the automatic type determination is a very intriguing feature, it was a
great suggestion! Infact it isn't too separate from task #13563 (No need for
polygon calculation when simple transformations are needed). Essentially, when
we have easy transformations (not just `scale', for example also flip, or
rotation by 90 degrees), we don't need to go through all the trouble that
ImageWarp goes in now to allow any kind of Homographic transform
<https://en.wikipedia.org/wiki/Homography> (calculating the area of each
output pixel on an input pixel and etc.). This will make the processing on the
simple transformations much more faster. In the process, we can also do
automatic type determinations like what you mentioned for the simple
transformations.

Based on your suggestion I thought of a more general solution that can apply
to any of the programs, not just ImageWarp. For example, I am really impressed
by how the `%g' option in C works: it will print a floating point number into
the most accurate representation (for example printing `1' for `1.0000000',
and the proper number of decimals for floating points). I would be really
curious if we could implement something like the `%g' and check the best type
for all the pixels in an image, and based on that, choose the best type for
the whole image without loosing mathematical precision. 

It doesn't just have to be applied to floats, so when the data-set is in
integer type, it will also check and for example if none of the values are
larger than `255' and `>=0', then use `unsigned char' and so on. We can do all
of this with a common option like `--automatictype' so when the user doesn't
like the default type of a program's output, they can add this option and get
the most appropriate type output. It can also be implemented as an operator in
Arithmetic.

But I am too busy to implement the `--automatictype' option right now, so if
anyone is interested, please post a comment here and start the work, I would
be happy to help ;-).

    _______________________________________________________

Reply to this item at:

  <http://savannah.gnu.org/task/?14362>

_______________________________________________
  Message sent via/by Savannah
  http://savannah.gnu.org/