#pdfparser

Python binding for libpoppler - focused on text extraction from PDF documents.

Intended as an easy to use replacement for pdfminer, which provides much better performance (see below for short comparison) and is Python3 compatible.

See this article for some comparisons with pdfminer and other approaches.

Binding is written in cython.

Requires recent libpoppler >= 0.40 - so I'd recommend to build it from source to get latest library, but it works also with recent libpoppler library present in common linux distributions (then it requires dev package to build). See below for installation instructions.

Available under GPL v3 or any later version license (libpoppler is also GPL).

How to install

Below or some instructions to install this package

Debian like - system wide libpoppler

sudo apt-get update
sudo apt-get install -y libpoppler-private-dev libpoppler-cpp-dev
pip install cython
pip install git+https://github.com/nghiapq77/pdfparser

Mac OS

pip install cython
pip install git+https://github.com/nghiapq77/pdfparser

Speed comparisons

pdfparser code used in test

import pdfparser.poppler as pdf
import sys

d=pdf.Document(sys.argv[1])

print('No of pages', d.no_of_pages)
for p in d:
    print('Page', p.page_no, 'size =', p.size)
    for f in p:
        print(' '*1,'Flow')
        for b in f:
            print(' '*2,'Block', 'bbox=', b.bbox.as_tuple())
            for l in b:
                print(' '*3, l.text.encode('UTF-8'), '(%0.2f, %0.2f, %0.2f, %0.2f)'% l.bbox.as_tuple())
                #assert l.char_fonts.comp_ratio < 1.0
                for i in range(len(l.text)):
                    print(l.text[i].encode('UTF-8'), '(%0.2f, %0.2f, %0.2f, %0.2f)'% l.char_bboxes[i].as_tuple(),\
                        l.char_fonts[i].name, l.char_fonts[i].size, l.char_fonts[i].color,)
                print()

How to modify parsing algorithm?

As you probably know PDF is document format intended for printing, so all logical structure of the text is lost (paragraphs, columns, tables, etc.). libpoppler is trying to reconstruct some of this logical structure of the document back by comparing physical positions of characters on the page and their mutual distances and reconstructing back words, lines, paragraphs, columns.

Component which is responsible for this reconstruction is C++ class TextOutputDev (in poppler/TextOutputDev.cc). It's using many constants for this jobs, vast majority of constants in hardcoded into code. Actually the only parameter that is available to Python code is combination of parameters phys_layout and fixed_pitch, which influences how text is ordered into columns. If you put phys_layout to True and fixed_pitch to value > 0, then fixed_pitch will be used as maximum distance between words in a line and minimum distance between columns (in pixels). I think phys_layout also influences order of boxes in page iteration. However influence of these parameters is not quite straight forward - so you'll need to experiment to see how it works in your case.

Another problem I encoutered is vertical spacing between lines in single box (paragraph) - this parameter is unfortunatelly fixed in libpoppler - it's constant maxLineSpacingDelta in poppler/TextOutputDev.cc, which is set to 1.5 (font size). If you need to accept bigger line spacing in paragraph then, you'll have to change it in C++ code and recompile libpoppler (in this case I recommend to make library local to pdfparser package). I've tried with value 2.0 and it seems to work fine.