The aim of this repo is to compare some tools to process pdf into text. The tools are :
- pdfplumber : made by jsvine for detailed information about each char, rectangle, line, et cetera — and easily extract text and tables.
- pdftotext : A simple PDF text extraction.
- pymupdf: a high-performance Python library for data extraction, analysis, conversion & manipulation of PDF (and other) documents.
- pypdf2: A pure-python PDF library capable of splitting, merging, cropping, and transforming PDF files.
- unstructured: an Open source libraries and APIs to build custom preprocessing pipelines for labeling, training, or production machine learning pipelines.
To compare the libraries, we used 3 PDF files:
multicolunm-code-travail-cmr.pdf: A multicolumn file containing the Cameroonian Labor Codewtables-benin-LF-2023.pdf: A conventional PDF with some tableswtables-food-calories.pdf: Another conventional PDF with some tables
For each library we extract the text from all files. Then give an annotation on a scale of 1 to 5 for each comparison criteria.
total time: Total time to process all the files;Conventional text: How well the tool performs on conventional textMulticolumn: How well the tool performs on multicolumn pdf documentTable: How well the tool process simple tables (each table line on 1 document) in pdf documentsComplex Table: How well the tool process complex tables (some table lines are on multiple document lines)Consistancy: How well the tool is viable (performs the same good result on different documents)
Following, is the comparison table of the results
| library | total time | Conventional Text | Multicolumn | Table | Complex table | consistancy |
|---|---|---|---|---|---|---|
| pdfplumber | 18s | 3.5 | 1 | 2.5 | 2 | 2 |
| pdftotext | ||||||
| pymupdf | 0.4s | 4 | 5 | 4 | 3 | 4 |
| pypdf2 | 7.8s | 4.5 | 5 | 4 | 2.75 | 4 |
| unstructured | 17.7s | 1 | 4 | 1 | 2 | 1 |