Size and alignment of a digit in an image for neural network training

Hi, I have been working on OCR using the Neural Network toolbox but I have a few fundamental doubts. Suppose I have the following images of digits:--

Since the digits are of different sizes, will it affect the neural network training and simulation? What are the parameters that a neural network uses to recognise images(like size, pixel values, shape, alignment, etc)? Please advise.

Answer 1

[Greg Heath]

The input to the net from a single image is just a column vector of numbers of dimension I obtained by feature extraction. The feature extraction of an image of size r x c can be complicated or as simple as columnisation via

input = image(:);

size(input) = [ I 1 ] = [ r*c 1 ]

Hope this helps.

Thank you for formally accepting my answer

Greg

[Soumyadip]

yeah, I know that. What I am asking is that if I train my network with the image of the bigger "7" above and then simulate it with the smaller "7", will the network recognise the smaller "7" as being similiar to the bigger "7"?

[Greg Heath]

It depends on your feature extraction. The net can only learn what you teach it.

If you use the simplest form, i.e., columnisation, it probably will not work unless you train on a range of sizes.

If your feature extraction involves size normalization then it probably will work.

[Soumyadip]

Thanks a lot. Can you provide me any resource on this feature extraction and size normalization stuff?

[Greg Heath]

No. However, maybe one of the resident image experts will respond.

If there is no response, try making that another posted question.