Analysis of a biologically motivated neural network for character recognition

A neural network architecture for size-invariant and local shape-invariant digit recognition has been developed. The network is based on known biological data on the structure of vertebrate vision but is implemented using more conventional numerical methods for image feature extraction and pattern classification. The input receptor field structure of the network uses Gabor function feature selection. The classification section of the network uses back-propagation. Using these features as neurode inputs, an implementation of back-propagation on a serial machine achieved 100% accuracy when trained and tested on a single font size and style while classifying at a rate of 2 ms per character. Taking the same trained network, recognition greater than 99.9% accuracy was achieved when tested with digits of different font sizes. A network trained on multiple font styles when tested achieved greater than 99.9% accuracy and, when tested with digits of different font sizes, achieved greater than 99.8% accuracy. These networks, trained only with good quality prototypes, recognized images degraded with 15% random noise with an accuracy of 89%. In addition to raw recognition results, a study was conducted where activation distributions of correct responses from the network were compared against activation distributions of incorrect responses. By establishing a threshold between these two distributions, a reject mechanism was developed to minimize substitutional errors. This allowed substitutional errors on images degraded with 10% random noise to be reduced from 2.08% to 0.25%.