A comparison of pre-processing transforms for neural network classification of character images

Optical character recognition (OCR) is the automatic conversion of textual information on a physical medium to a useful electronic form. In common with most visual pattern recognition tasks it is commonly divided into three sub-tasks; image capture, image processing and image classification. This work is based on the conjecture that the combination of unitary image transforms, an image processing technique, with neural network classifiers is a potentially useful method for performing OCR. The investigative work consisted of a series of experiments that involved taking live character image data, performing a set of unitary integral transforms on that data, and evaluating the effect of the transforms on the learning behaviour and recognition performance of a fully-connected, three-layer, backpropagation network. This was done by comparison with a network trained on the un-preprocessed image data