parsee-pdf-reader 0.1.8.2

No project description provided

Parsee PDF Reader

This PDF reader was designed to overcome the common problems when trying to extract tables from PDFs.

We started initially with a focus on financial/numeric tables, so this is what this PDF reader works best for.

This is an early release, so we will be still making major changes.

Installation

Recommended install with poetry: https://python-poetry.org/docs/

poetry add parsee-pdf-reader

Alternatively:

pip install parsee-pdf-reader

In order to use the OCR capabilities you also have to install tesseract: Install Google Tesseract OCR (additional info how to install the engine on Linux, Mac OSX and Windows). You must be able to invoke the tesseract command as 'tesseract'. Note: in our testing we always used tesseract 5+, as that proved to be the most reliable. So for Linux you might have to build from source to get tesseract 5.

In order to run the PDF to image functionality you need to install poppler, e.g. on MacOSX:

brew install poppler

Extracting Tables and Paragraphs

Extracting tables and paragraphs of text can be done in one line:

from pdf_reader import get_elements_from_pdf
elements = get_elements_from_pdf("FILE_PATH")

If you are processing a PDF that needs OCR but no elements or just very few are being returned, you can force OCR like this (replace the paths):

elements = get_elements_from_pdf("FILE_PATH", force_ocr=True)

If you want to visualise the output from the extraction, you can run the following (replace the paths):

from pdf_reader import visualise_pdf_output
visualise_pdf_output("FILE_PATH", "OUTPUT_PATH")

This will save an image of each page with the detected tables and text marked in red.

Methodology

Combines pdfminer, pypdf and tesseract and augments them with the introduction of table elements, which are treated separately from the rest of the text. As a result, the output contains basically two types of elements: tables and text paragraphs. We believe this separation is important as otherwise the tabular information is not extracted very precisely and concepts such as columns and rows are usually lost.