How to stamp OCR layer back on original PDF?

[gboor]

I have a scanned PDF with some text and images in it. If I go page by page and use page.get_pixmap().pdfocr_tobytes(...) I can convert all pages into OCRed pages and recombine those back into a single PDF - so far so good.

However, the images get pretty garbled in the process. The resolution is messed up and the sharpness is way down. I know I can probably mess around with the DPI, but ideally I would not generate a new PDF here; I would like to take the text layer and re-insert it into the original PDF and not mess with the images at all.

I've experimented with get_textpage_ocr(...).extractBLOCKS() and then inserting those with insert_textbox, but it seems the coordinates are way off and the text size isn't right either.

Is there any way to just insert a new transparent text layer from OCR into the original PDF?

[JorjMcKie]

Try this:

pix = page.get_pixmap(dpi=<decent-value>)
pdfdata = pix.pdfocr_tobytes()
ocrdoc = pymupdf.open("pdf", pdfdata)
ocrpage = ocrdoc[0]
ocrpage.add_redact_annot(ocrpage.rect)
ocrpage.apply_redactions(imges=pymupdf.PDF_REDACT_IMAGE_REMOVE,
    graphics=pymupdf.PDF_REDACT_LINE_ART_NONE,
    text=pymupdf.PDF_REDACT_TEXT_NONE,
)
# only text is now left over
# put it underneath original page
page.show_pdf_page(page.rect, ocrdoc, 0, overlay=False)

You can choose any DPI for Pixmap creation, because show_pdf_page() will scale to fit on original page.

[JorjMcKie]

Writing generated OCR text back into original bboxes is a tricky thing:

• Tesseract's internal font ("GlyphLessFont") is monospaced. You have to decide for Courier therefore.
• The detected bboxes (not necessarily the text itself) is often ludicrous.
• There exists no font size info - all you have are those bboxes, from which you must compute a meaningful value for font size (which goes wrong in too many cases consequently).
• Method .insert_textbox() is consequently also often enough just wrong.

[gboor]

The redaction pathway seems to work. The text is in the right place now.

It still feels "hacky". I previously used to do this with tesseract-ocr on the commandline (no longer an option, hence the switch) by generating hOCR output and splicing that back into the PDF calling cpdf (also not an option anymore).

It doesn't look like pymupdf has support for tesseract's hOCR functions, though and even that required going through a bunch of additional tools to make it work neatly.

This works very well using just one dependency and no external binaries, so I'm happy :)

[JorjMcKie]

It still feels "hacky".

Well, that's a matter of taste ... or even mannerism!
Re-inserting the detected text in to the original page by .insert_textbox(...) ... that is hacking. And even what OCR engines do, going by your strict standards, that would qualify as ‘hacking’ even sooner.

The result of pix.pdfocr_tobytes() creates a 1-page OCRed PDF. We could have simply replaced the original page by that page - however, its width / height dimensions are a consequence of the DPI decision and thus not equal to the original.
If this is of no concern, then voilá: could be done just as well.
Or we could have emptied the original page completely (image(s) and whatever else on it) with a redaction annotation and copied the OCR result as a whole onto it using show_pdf_page().

I think overall this is a clean approach with minimal intrusion ...

But of course I am happy, that it solves your problem.

[gboor]

I didn't mean that in a negative way and you are correct on all points, of course.

I guess any approach here is going to have a fair bit of "hackiness" simply because of the nature of the problem and the nature of PDFs and OCR not natively gelling too well.

[Vencent-art]

Redact PDF tools：https://redactionpdf.com/redact-pdf-online