The way I tried to approach this, was using pytesseract and OpenCV, but it’s not perfect.
The goal is pretty much something where I take a picture of a receipt and automatically add it to an app which will nicely depict spendings.
First of all I guess a more important question is: Is there a good place to look for advice on this, apart from google and this forum? If so, please let me know.
So back to the details: For now, in order to have it work reliably I would probably have to implement some sort of manual check, in order to correct the mistakes the program made, but here’s the thing; All of these receipts are pretty much the same, so I am hoping to find some way to show the program a couple of examples, as well as the correct output and have it perform more reliably.
So how could I approach this?
Edit: Another thing that I am wondering about is, how exactly I could create an App, that would pretty much automatically send files (pictures of the receipts) to my Computer, or at least give the computer access to them.
Google and stack overflow, I guess. The top Google results look like they are copies of the same tutorial essentially. Something like this will require some research into the hopefully wonderful tesseract and opencv documentation, much like the tensorflow projects in the data science curriculum. And as you mention, OCR isn’t perfect and even with the AI improvements, I would be surprised if hundreds or thousands of images were not needed to get reliability. The verification step you suggest would probably be necessary regardless.
This will be the easy part. If you are using python, just create a django or flask web app and upload the picture for processing.
Good luck. I’m interested in this generally as I’ve had several similar ideas for processing text like this, so I’ll be interested to see how this works out for you.
Thanks for the response!
Yeah, I was hoping to eventually create an option of “Trust the AI” that could disable the manual verification step. I don’t really know too much about Machine Learning, but I guess I will try to make it work as it is right now and perhaps try to combine the current solution with some sort of personal text detection later on.
But again, no idea how hard that might eventually get, since text recognition seems to be something, where technology is not yet as advanced as one might expect/hope.
Currently I am still experimenting what kind of picture (file type, filters,etc.) work best, but it seems to do a more or less decent job of recognizing most of the important parts.
I will let you know if I make any significant process/when I have finished the project.