PDF (magazine style columns) to ePub Conversion

[therealjoeblow]

First, I know that PDF is a bad source to try to convert from. It's the only source I have though.

The PDF is laid out "magazine style" with the text in columns, meant to be read down column 1, then down column 2:

This is some text               The text in the second
that is in the first            column is formatted like
column.                         this.  I guess you get 
                                the idea.
It has paragraphs
laid out like this.

...And when I convert to ePub, I end up with lines that scramble the paragraphs from column 1 and 2 together making it impossible to read comprehensively:

This is some text
The text in the second
that is in the first
column is formatted like
column.
this. I guess you get
the idea.
It has paragraphs
laid out like this.

...Basically every 2nd line belongs with a different paragraph from another column.

Is there any way to convert this successfully?

Cheers
TRJB

[theducks]

Some better OCR programs can be configured to deal (mostly) with 2 columns.
Those still need hand holding to determine which SHOULD join the article and which belong elsewhere (eg continued from page n)

[kovidgoyal]

If you have microsoft word, try importing it into that. It has a more spohisticated analysis engine that handles multi-col PDFs sometimes.

[therealjoeblow]

I just <ctrl-a> copied all of the PDF, and <ctrl-c> pasted it into a .txt file

All of the text copied over in the correct order, then I just used regex in notepad++ to clean up line breaks like I would have to anyway, and then made a epub with calibre; then just manually inserted the graphics where they belonged.

A bit tedious, but got it done.

Although that makes me wonder... if simple <crtl-a> is able to copy all of the text in the correct order, why is Calibre not able to process it in that way?

That part doesn't make sense since simple 'copy all' is probably one of the most rudimentary text tools and it seems to be able to handle the columns.

Cheers
TRJB

[kovidgoyal]

calibre does not extract text from PDFs, that is done by pdftohtml from the poppler project, feel free to ask them why it does not do a better job.

[therealjoeblow]

Ok, thanks.

It wasn't meant as a criticism of your awesome work BTW!

Cheers
TRJB

[kovidgoyal]

Yeah, it's been on my TODO list forever to get rid of pdftohtml and write my own text extraction engine, but...

[retiredbiker]

Some better OCR programs can be configured to deal (mostly) with 2 columns.
Those still need hand holding to determine which SHOULD join the article and which belong elsewhere (eg continued from page n)

I use the free Tesseract OCR that comes with Ubuntu, and the OCRfeeder GUI front end. Most times, it will handle double columns correctly, but not always. And if there are embedded pictures, breaks with images or huge caps, and son on, I often have to do smaller areas one at a time. Also for the "continued on..." case. Can be tedious, but it works pretty well, and is amazingly accurate in detecting paragraph breaks and dealing with end-of-line hyphens. I do it page-by-page, proofing as I go, so I'm not faced with a massive proofread task at the end.

As for the copy/paste technique, I find that all depends on what is in the individual PDF. Some work fine, some are simply impossible. If the text layer is just junk, OCR is the easier answer.