To extract all tables in PDF

[wingbroken]

I know there are some examples that can extract table contents from PDF with Pymupdf. But seems they require pre-analysis of specific doc to adress the table. Any generic way we can extract all tables in PDF like pdfplumber.extract_table() does via some prior configuration? If not, consider to enhance Pymupdf? I like PYMUPDF's fast speed and more granular funtions.

[JorjMcKie]

I understand your motivation.
There are at least two or three considerations that let me turn down this enhancement proposal:

1. When exporting spreadsheet documents to PDFs (like from Excel or LibreOffice), the result will not conform to the table formalism as defined in the PDF spec. So an important (if not critical) condition of implementing this is not met.
2. Analyzing a table that is indeed specified using PDF spec, is a very, very complex undertaking.
3. Detecting and then analyzing tables, which are basically defined as text, is even more challenging. I have seen a few heuristics approaches which use pretty heavy calibers. Clearly beyond of what I consider PyMuPDF's scope.

Taking the three together, there is no business case for me doing this.
As PyMuPDF does offer ways accessing low-level structures in PDFs using Python, nothing stands in the way for someone who wants to try it.

[ram02z]

3. Detecting and then analyzing tables, which are basically defined as text, is even more challenging. I have seen a few heuristics approaches which use pretty heavy calibers

Could you please point me to one of the approaches you mention?

[JorjMcKie]

For example here: https://medium.com/mlearning-ai/extracting-tabular-data-from-pdf-using-deep-learning-table-detection-81242cbad6d

But also this: https://www.pdftron.com/pdf-tools/pdf-table-recognition/

[ram02z]

For example here: https://medium.com/mlearning-ai/extracting-tabular-data-from-pdf-using-deep-learning-table-detection-81242cbad6d

But also this: https://www.pdftron.com/pdf-tools/pdf-table-recognition/

Thank you!

[ram02z]

Sorry for the followup question, but do you know of any approaches that didn't use AI?