I have scanned documents in large pdf files consisting of many individual documents. Each document begins with a exhibit number sticker, much like this one: example. The files are scanned in greyscale.

I am looking to detect this numbered sticker object in the files. On top of that, my current understanding is that I would need to crop this sticker if I want to run another NN on it to recognize the handwritten digits on the sticker. I wish to split the files on the pages where an object is detected and rename those individual files based on the handwritten digit on the sticker.

So far I know I must split the pdf pages into individual images with pdf2image like in this example, and I should label the training data beforehand with something like labelimg. After reading more into it, it looks like tensorflow and keras will suit my project fine.

My questions are as follows:

  1. Assuming the image detection model generates a valid bounding box, how can I crop the bounding box?

  2. What NN would best suit each need (detecting single object + handwriting OCR)

  3. What output file should I expect from the model? Should I expect any obstacles in using the output file as a basis to process the corresponding pdf files with something like PyPDF2?

The short of it is I'd want to do what's described in this article and edit my pdf file based on the results.



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.