Exporting with alpha/transparency in a PDF

[jdonley]

I am losing transparency in a PDF file for areas that have alpha. MATLAB exports the PDF with the vector components displayed correctly with transparency but export_fig does not.
This sounds similar to issue #3 but I am pretty sure MATLAB isn't embedding a rastered version in the PDF (I can zoom all the way in and the shaded areas are still sharp).

[altmany]

What are the two commands that you're using (export_fig and print)? in other words, which parameters are you using? Also, which platform and Matlab release?

[jdonley]

MATLAB R2015a 64bit
print('-dpdf', 'blah_matlab.pdf');
export_fig('blah.pdf');

Also doesn't work with '-transparent' flag, but that is expected according to the docs for export_fig

[jdonley]

Oh, I see... export_fig saves in EPS then converts back to PDF.
EPS doesn't support transparency but can we maintain the transparency though the conversion using pdfmark?

[altmany]

I am not familiar with pdfmark, but if you can modify eps2pdf.m in a way that would fix the problem then I would be happy to incorporate your fix in the main version.

[jdonley]

I am working on a workaround for this.
I need to add the line:
{alpha} .setopacityalpha true
to the eps file with the appropriate alpha value in replacement of {alpha} whenever there is a transparent object.
Any ideas on how I can locate only the elements which had transparency in the EPS file?

[altmany]

I think it would be best to create a super-simple white figure that only has a single simple semi-transparent alpha-patch, and then generate the EPS export using the following command:

print(gcf, '-painters', '-loose', '-depsc2', 'test.eps')

This would make the EPS text file much smaller and you should be able to easily identify the alpha patch near the bottom of the file. EPS files are simply text files that contain PostScript directives, so you can easily experiment with your solution by simply editing the generated EPS file in any text editor (e.g., the Matlab editor) and then re-displaying the image file to see if the change helped.

Once you figure out how to post-process the EPS manually, this can be automated. You can look at print2eps.m (below line #233) for some EPS post-processing examples. The basic idea is to use the regexprep Matlab function to replace an existing EPS (=PostScript) directive with a corrected one.

Let me know if you still need my additional help with this.

[jdonley]

I have already done pretty much exactly what you've suggested haha.
What I mentioned before is exactly what needs to be done and it works (i have tested).
I just can't isolate the patches in the EPS that will require transparency, that is the tricky bit.
The information is lost when we go to EPS. I need to know which patch objects in the EPS originally had an alpha component in the matlab figure.

This works for making all patches transparent:
fstrm = strrep(fstrm, sprintf('L\ncp\nf\nGR'), sprintf('L\ncp\n.5 .setopacityalpha true\nf\nGR'));

[altmany]

Suggestion: take a look at the color values. If you don't see anything obvious, you could possibly modify them as a signal. For example:

1. save the original colors of all alpha objects in the fig
2. then modify all alpha patches to have a color of [101,102,id]/255 where id is a numeric sequence id (1 for the 1st patch, 2 for the 2nd etc.). This assumes that a [101,102,*] color is not used by any handle in the figure; otherwise, you can dynamically select any different static RG-duo that is not in use.
3. then generate the EPS using print
4. then post-process the EPS file and replace all patches that have an RGB of [101,102,id] with the real RGB value for each patch id, and its appropriate alpha directive
5. finally, remember to restore the fig's patches back to their original RGBs

[jdonley]

This shows the fix (#111) (notice the thin lines from export_fig compared to the thick lines from MATLAB (yet still the thinnest MATLAB can do) )

Fix may get stuck on rare cases where ID color values are the exact same as other patches but should be good for most of the time.

[altmany]

Fixed via merge of pull #111