A new training rule for optical recognition of binary character images by spatial correlation

Explores a method for automatically training and recognizing patterns such as characters or symbols in an image. The research is based upon the commercially proven recognition technique of spatial correlation, whose major drawback is the tedious process of training each character while taking into account the variations in print from sample to sample. The research attempts to completely automate the training process by a new learning rule in feedforward neural networks, to create an “optimal” representation of each character from a representative set of character images. The research presents bounds of the learning constant and proofs of convergence of the proposed algorithm. The method significantly enhances existing commercial OCR systems that are based on spatial correlation