Position Invariant Optical Character Recognition through Symmetry Features

We propose an effective method to achieve position invariance in the application of optical character recognition (OCR). We normalise the position of all inputs based on their symmetry features. The generalized symmetry transform (GST) is used to determine the symmetry features prior to classification by a probabilistic neural network (PNN). We used the United States Postal Service (USPS) data set to measure performance.