I have not done plate recognition, and have only dabbled in neural networks, but I would expect that you need to translate your segmented characters into a vector. That means you first need to standardise the width and height of each segment, which will involve resampling. I would also expect some kind of image normalisation.
Then you can supply each segment as a 1-dimensional vector. The number of pixels in that vector is the number of inputs in your neural network. From memory, the neural network expects inputs along the rows, and each individual training case in the columns. So you would insert each segment-vector into a unique column of your input matrix. The corresponding training output vector, of course, is the character that each segment-vector represents.
Something like that, anyway =)
