Ghostscript is more reliable than Poppler's pdftoppm renderer. gs is
also a hard dependency, as the only open source tool that can produce
a PDF/A file, while Poppler could be removed. pdftoppm has awkward
syntax with some special handling needed for different versions. I have
found isolated rendering bugs with pdftoppm as well.
With that, I'm removing supporting for multiple rasterizers.
A minor advantage of pdftoppm is that its code produced JPEGs where
possible, but this can be achieved with gs.
The immediate reason for doing this is that (newer?) versions of parse()
seem to choke on the parse string. It appears to trigger exponential
behavior in the underlying regex. In any case, replacing subprocesses
with native Python is usually better.
Prior to this change, hocrtransform would render printable text (black
on white) and then a fully opaque image on top of the text. According to
the PDF spec, text that is the output of OCR should be marked invisible,
so that PDF viewers /know/ it's OCR output in a document that might mix
OCR and text overlays. Another benefit is that PDF viewers would know
to skip rendering text if they are not smart enough to figure out the
image will completely overwrite it.
However, for debug, visible text is nice, so retain it as an option.
As documented, Tesseract does not escape the filename when inserting it
into .hocr, potentially creating an invalid XML file as a result. Since
there is no use for the title, regex it and nuke it.
Of course, this introduces recompression artifacts, and is unnecessary
if no options are given that modify the final image (no -d, -c, -i).
But rather than worry about that, it would be better to ultimately find
a way to combine the original PDF page with the output PDF text in the
case where we want no changes to the original. This is good enough for
now.
The better option can apparently be achieved using pdftk background, or
probably better, PyPDF2's merge. If Tesseract PDF generation is used
then we need a way to remove the image. Tesseract PDF generation at 3.03
does layout better (I think) and also properly encodes the hidden layer,
which is less likely to give display issues (I think).
Ghostscript has the clunkiest imaginable syntax, obtuse documentation,
quirky behavior, and poor diagnostics... but it *actually works* unlike
pdftoppm/poppler which gets things wrong.
In this case I observed poppler incorrectly decompresses certain CCITT
encoded monochrome PDFs. So set up Ghostscript to do the job instead.
For the moment this performs monochrome -> RGB conversion via reportlab.
It appears to be possible to have a PDF with an embedded font that is
either unused or used only for whitespace. So check for some amount of
actual text instead.
Jim Barlow