Pfanzagl (1982) is a classic text in semiparametric statistics but is difficult to read due to its archaic typesetting. This repo processes the text using OCR (via gemini 2.0) and converts it into latex, which can then be post-processed into a reasonable output format.
This can be adapted to OCR and typeset your own old books/papers. The process is as follows:
- Scan the text using a high-resolution scanner (300 dpi or higher), preferably in grayscale, output to pdf.
- Use
0_extract.pyto convert the pdf into images (this populates thepdf_imagesdirectory).
- python requirements in
requirements.txt - also requires GhostScript installed (e.g. via
brew install ghostscripton MacOS, orapt install ghostscripton Ubuntu)
- Use
1_ocr.pyto run OCR on the images (this populates theraw_texdirectory).
- uses
google-genaito call gemini 2.0. RequiresGEMINI_API_KEYin your environment variables. You can get this from google ai studio
- Tinker with the output in
raw_texto get the text into a readable format. This is the most time-consuming part of the process, but is hard to automate. An example of a successful output is shown below, which is generated from page 51 of the text and is1_tex_paper.texin theraw_texdirectory.