OCR with PyMuPDF

[ghost]

I noticed that when calling page.get_text it extracts text even when PDF is image based, is Tesseract used to do that?

Is there a way to specify a document language and use corresponding OCR config?

[JorjMcKie]

If text can be extracted, then that text is already present in the file. It is not dynamically generated by whatever mechanism.
So, yes, some OCR tool has probably been used to put an (invisible/hidden) text level underneath the page images and the resulting file has been so modified.
Only during that OCR process the language to interpret the images' text can be selected. This is btw just an interpretation help to speed up the OCR engine's work.
When your PyMuPDF script extracts text, then all that has already happened and is set in stone.

[JorjMcKie]

While the base library MuPDF offers integrated use of Tesseract - thus providing a somewhat "dynamic" OCR feature - this is not (yet) supported by PyMuPDF.
But you can install OCRmyPDF, import it in your Python script and invoke it page-by-page using PyMuPDF - resulting in a similar behaviour.
The basic approach would be to make a 1-page PDF, pass that to ocrmypdf, receive back that temp PDF with its new text layer and then extract the text.
While this does work in principle, I haven't yet a ready-to-go code snippet ...

[ghost]

Thanks

[JorjMcKie]

I am about to publish a little script that does what I described Looks like so:

"""
This is a basic script demonstrating the use of OCRmyPDF together with PyMuPDF.

It reads a PDF's pages and passes them to ocrmypdf one by one. One could at this
point insert some checks as to whether the page is actually an, contains no text,
or text with many unrecognized characters or the like.

Each page is then converted to a 1-page temporary PDF which is
- passed to ocrmypdf for OCR-ing it
- the 1-page output PDF of the pervious step is then text-extracted
- return the extracted text

Instead of extracting simple naive text format, one could also use all other
text extraction formats like "dict" to get text position information.
"""
import fitz
import ocrmypdf
import sys
import io


def ocr_the_page(page):
    """Extract the text from passed-in PDF page."""
    src = page.parent  # the page's document
    doc = fitz.open()  # make temporary 1-pager
    doc.insert_pdf(src, from_page=page.number, to_page=page.number)
    pdfbytes = doc.tobytes()
    inbytes = io.BytesIO(pdfbytes)  # transform to BytesIO object
    outbytes = io.BytesIO()  # let ocrmypdf store its result pdf here
    ocrmypdf.ocr(
        inbytes,  # input 1-pager
        outbytes,  # ouput 1-pager
        language="eng",  # modify as required
        output_type="pdf",  # only need simple PDF format
        # add more paramneters, e.g. to enforce OCR-ing, etc.
    )
    ocr_pdf = fitz.open("pdf", outbytes.getvalue())  # read output as fitz PDF
    text = ocr_pdf[0].get_text()  # ...and extract text from the page
    return text  # return it


if __name__ == "__main__":
    doc = fitz.open(sys.argv[1])
    for page in doc:
        text = ocr_the_page(page)
        print("Text from page %i:" % page.number)
        print(text)

It already works under Windows. Now testing Linux with it.

[JorjMcKie]

Confirming: Linux also works!

[JorjMcKie]

You can try it out here: https://github.com/pymupdf/PyMuPDF-Utilities/tree/master/OCR