The purpose of this project, was to create application able to recognize computer encoded character, basing on its image in .tiff format. The application uses Hopfield network, designed to store and recognize characters patterns. Although the implementation of the network and application design were successful, results are not satisfying. A single Hopfield network, made of 120 neurons is too small to store all letter patterns. In consequence, a flow the of energy function is full of noise from overlapping patterns stored in the network and the output is often false local minimum. Unfortunately the network is not capable of recognizing patterns used in learning process, therefore using it to recognize noisy patterns or handwriting would not be effective. As shown by the test, recognition capabilities grows as the number of training patterns decrease.
Better results could be achieved by applying bias to energy function, in order to more proficiently find the solution. Alternatively network could be dispersed on few smaller networks, where each of them would process only a part of the image and the result would be calculated based on sum of their outputs. On the other hand, Hopfield network in this form was originally presented in 1982. Modern methods, based on convolutional or deep networks, achieves far better results in the field of Optical Character Recognition.