At the forefront of Artificial Intelligence

Home » Articles » Neural Networks » Projects

Simple OCR Project

Create a program that can recognize characters given an "image" consisting of text.

Guidelines

The problem can be greatly simplified if the domain is restricted. To the right, you can see the sort of circumstances that a typical OCR system might have to deal with. From top to bottom: noisy data, tilted data (from scanning), different fonts, different sizes.

Now, if you limit your OCR system to be able to deal with only the first case (noisy data) and assume the characters are always perfectly horizontal and they are of a consistent font and size. It makes the problem much easier, while remaining a challenge.

OCR can be done through several methods: pattern matching, statistical analysis (essentially, still pattern matching!), or perhaps neural networks (Generation5's ONR program uses perceptrons to recognize numbers).

If you unfamiliar with programming graphical interfaces (or reading from image files) you can use text-based images just as well. Again, ONR works like this - reading from two files, one with the numbers 1-9 which it then learns from. The other data file is a series of "corrupt" numbers that ONR tries to recognize.

Solutions

ONR	A Generation5 example program. ONR uses perceptrons to learns the digits 1-9, and then recognizes corrupts versions of the digits.

Submitted: 14/04/2001

Article Toolbar

BibTeX entry