Saving the images ocrmypdf temporarily creates OR use existing pdf-to-img pdfs

[jaspermowatt]

Hi everyone,

I currently have an OCR workflow that uses Surya OCR for ocring and table recognition and generating text files and CSVs from scanned semi-technical PDF documents. These outputs are then fed into an LLM data extraction tool. While Surya performs well, it doesn't create searchable PDFs. To address this, I run OCRmyPDF on the original PDF as a final step to generate a searchable PDF.

However, I've noticed an inefficiency in this approach since both Surya OCR and OCRmyPDF perform PDF to image conversion and preprocessing. To optimize the workflow, I'm looking for a way to either:

a) Provide the images generated by Surya OCR as input to OCRmyPDF, avoiding the need for OCRmyPDF to convert the PDF to images;
OR
b) Change the order of my workflow to run OCRmyPDF first, retaining the images it creates during its pipeline. Then, modify the Surya OCR workflow to use these pre-generated images as input.

The goal is to perform the PDF to image conversion only once throughout the entire workflow.

If anyone has experience with modifying OCRmyPDF to achieve either option (a) or (b), or has an alternative suggestion for streamlining this process, I would greatly appreciate your insights and advice.
Thank you in advance for your help!

Best regards,

Jasper

[jbarlow83]

You could create a OCRmyPDF plugin that uses Surya as its OCR engine instead of Tesseract for example.
There's an OCRmyPDF-EasyOCR that demonstrates how this could be done (although the more current approach would be to render to hOCR).

There are often conflicting goals in the latest LLM OCR compared to conventional OCR; in particularly they're strongly focused on accuracy, which is great, but not on pixel-perfect text positioning or grouping text into paragraphs. In particular, Surya only obtains text line level positioning, so for typical variable width fonts you'd see significant misalignment if it were used to generate PDF.

You could also have your OCR plugin run both Surya and Tesseract OCR, saving Tesseract for PDF generation and setting aside Surya elsewhere.

[jaspermowatt]

Thank you for the fast response. I think I understand you, so this is why I recognise using surya to generate the searchable pdf is not an option and why I would like to use surya to generate my text files and table files, but then use tesseract-OCRmyPDF to generate the pdf. So basically the route suggested in your last point.

Apologies, perhaps I did not explain well enough, my aim is not to use surya to generate a searchable pdf, but rather reuse the images created in the pdftoimg step that i believe is common to both surya and OCRmyPDF.
I've attempted to draw a diagram of the ideal workflow that might explain things better:

The surya output (jxt/json and CSV) will be used in any case for my llm based data extraction tool whilst the OCRmyPDF output (searchable pdf) is simply so I have a searchable/highlightable pdf at the end of the process.

Right now my workflow comprises two separate processes but in both situations, they are taking the input pdf and converting it to images. This is the longest step, so I'd like to use either my surya workflow or my OCRmyPDF workflow to create the images, and then the subsequent OCR step simply uses those already generated images, rather than repeating this step.

Thank you