benjaminpstern / optical_recognition Goto Github PK

Optical Character Recognition project

This is a project I'm doing to test a new algorithm for recognizing characters:

suppose you have an image of a character, how do you represent it? Most programs either divide the pixels into grid sections and store the average darkness of each section or try to identify features in the character.

This algorithm seeks to use a representation that correlates more closely with the features we see as humans.

The idea is that if we draw a line from the center of mass of an o along any direction(making an angle theta with the horizontal), it will intersect the o an equal distance away no matter how you change theta.
But if we do this for a b, we will intersect the stem at an angle around 120 degrees, as well as the loop.
If we do this for a c, we won't intersect between about -30 and +30 degrees. This information is fast to store, fast to find, and is much more handwriting-independent than pixel grids.



Requirements: pillow, numpy, scipy, matplotlib, scikit-learn on python2

To recognize a character, create a Picture object of it using the picture class and make a Char_Repr object using that picture. then use scikit-learn to train on a data set and recognize characters. For an example see linearSVCTest.py