Probably want to use a neural network for this.
Neural networks are rather simple in theory, see
Wikipedia[
^]. They have the same architecture as the brain.
You basically give it a number of input neurons (1 for each pixel), 1 or more hidden layers of X neurons, and then a number of output neurons (probably 1 for each digit in the numberplate). Each neuron is attached to every neuron in the previous and next layer, building a network. Illustrated in
another Wikipedia page[
^]. Each connection then has a weight, and each neuron has a transfer function.
All weights are randomly initialised. You then give a number (typically thousands for this kind of project) of training sets. You give it an image, and what the numberplate is. This will train the network to recognise the plate.
If your images are taken from a road security camera you might need 2 networks. 1 to recognise each of the cars and then trim down the image to only fit the car (or even just the front of it), and a 2nd to find and read the numberplate on each car. This will give much more accurate results.
This article[
^] has a working implementation of a neural network which can recognise written digits. You should be able to use this as a basis for your network.
Submit an article or tip