JerryMungo
Honorary Master
- Joined
- Jul 18, 2008
- Messages
- 37,637
- Reaction score
- 6,390
.
Last edited:
-
Join the MyBroadband community
South Africa’s biggest forum. Discuss, discover, and connect with thousands of members.
-
Question of the Day: What is your preferred Chinese car brand?
Using machine vision for recognizing areas of data on a page
- Thread starter JerryMungo
- Start date
WAslayer
Honorary Master
You want to use zonal OCR for that.. as for approach, I played around with a python based OCR a while back, that has the ability to be trained.. don't recall if it had zonal functionality.. whe I remember the name, I'll confirm..
WAslayer
Honorary Master
Yes, tesseract, that's the one..
Also, on the already built for you and push your document to an API front, Amazon Textract already can do all of this for you.. exception is the zonal part, where you can't tell it where to look, but it can tell you in which zone of the document it found the data..
WAslayer
Honorary Master
Cool..!
Hope you don't run across things like '/' in dates etc.. OCR freaks out with the special characters and I had a terrible time trying to get it to recognize them.. sadly could not and had to abandon the idea.. but, my use case was quite niche and the strings I needed it to recognize was not natural language, so that didn't help either..
DA-LION-619
Honorary Master
Machine-Readable Zones (MRZ)?
Find out what is natively available on the device, then create the forms such that you can detect each zone and pass the image into your OCR engine of choice.
MagNorthDigital
Expert Member
Tesseract would work a treat for this, specifically in something like react-native
Just be aware that whatever you are reading off the screen might change from one day to the next... aka style changes, dark mode, browser extentions or crazy browser changes.
Also, if you are reading first party data, is there not a more direct way? It's a lot of hassle for something that could be very fragile.
Just be aware that whatever you are reading off the screen might change from one day to the next... aka style changes, dark mode, browser extentions or crazy browser changes.
Also, if you are reading first party data, is there not a more direct way? It's a lot of hassle for something that could be very fragile.
MagNorthDigital
Expert Member
Oof, you have to deal with PBKAC too... If my experience with users has taught me anything, design the forms so that they fill in the letters in cells (like gov forms) and use black ink only.
If you have any control over the designer / form designs, ask / tell them to draft a high contrast version, and a "lighter grey" version (and then test both against your app). So that when the person filling the form uses black (or blue, cause you know...) you have as clear a reading surface as possible.
God speed good sir.
Ps. You're essentially trying to teach ocr to do this... https://www.myfonts.com/WhatTheFont/
Recognize letterforms
If you have any control over the designer / form designs, ask / tell them to draft a high contrast version, and a "lighter grey" version (and then test both against your app). So that when the person filling the form uses black (or blue, cause you know...) you have as clear a reading surface as possible.
God speed good sir.
Ps. You're essentially trying to teach ocr to do this... https://www.myfonts.com/WhatTheFont/
Recognize letterforms