Open benwbrum opened 6 years ago
I managed to segment out the document from the image and colour the the rough bounding boxes. After this preprocessing a Deep neural network algo ought to work given enough data.
In case there is not enough data then I can approach the problem through conventional Image processing techniques. I have a few ideas in mind.
I have implemented one such tool that returns bounding boxes for all fields in a form(of a bank) based on a given template using basic image processing tools and mathematical rules.Provided more data we could make it more efficient using deep learning.This should make for an interesting project
One of the obstacles our project (and similar projects) has encountered when trying to use existing online transcription tools is that most of these tools require users to identify the entry ("region of interest" or "annotation target") they are transcribing, drawing rectangles or lines on an image of the page. This is extremely frustrating to users, who prefer a mouse-free mode of data entry and do not see why the entries cannot be found programmatically.
We would like tools which locate each entry on a census form which needs to be transcribed. These tools should accept an image (or set of images) and produce sets of coordinates representing pixel locations of each entry on the image. These coordinates may draw rectangles around the entries, polygons around the lines of text, or single lines underneath the entry or through the middle of the entry.
Sample data: