R. A. Wilkinson

Methods for enhancing neural network handwritten character recognition

An efficient method for increasing the generalization capacity of neural character recognition is presented. The network uses a biologically inspired architecture for feature extraction and character classification. The numerical methods used are optimized for use on massively parallel array processors. The method for training set construction, when applied to handwritten digit recognition, yielded a writer-independent recognition rate of 92%. The activation strength produced by network recognition is an effective statistical confidence measure of the accuracy of recognition. A method of using the activation strength for reclassification is described which, when applied to handwritten digit recognition, reduced substitutional errors to 2.2%.<<ETX>>

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%.

Self-organizing neural network character recognition on a massively parallel computer

Two neural-network-based methods are combined to develop font-independent character recognition on a distributed array processor. Feature localization and noise reduction are achieved using least-squares optimized Gabor filtering. The filtered images are then presented to an ART-1-based learning algorithm which produces self-organizing sets of neural network weights used for character recognition. Implementation of these algorithms on a highly parallel computer with 1024 processors allows high-speed character recognition to be achieved in 8 ms/image with greater than 99% accuracy on machine print and 80% accuracy on unconstrained hand-printed characters

Using self-organizing recognition as a mechanism for rejecting segmentation errors

We have developed a method for self-organized neural network based segmentation error checking and character recognition. In this method the page is segmented, a pre-trained self- organizing classifier is used to classify the characters coming from the segmenter and these reclassified characters are used to adaptively learn the machine print font being segmented. This allows the self organizing character classifier to perform font and image quality checking while independently checking for segmentation errors. No context is used to correct segmentation or recognition errors since the page was randomly generated. In our experiments segmentation errors caused the sequential classification of segmented characters to be confused for 6.6% of the items segmented because of the splitting and merging of characters. By using self-organizing neural network classification 9.2% of these errors were corrected to produce a correct segmentation and classification rate of 93.4% overall. After self-organization correction, segmentation and classification was done to an accuracy of 99.3% with no human intervention and in all cases 99% of all segmentation errors were detected.

Massively parallel implementation of character recognition systems

A massively parallel character recognition system has been implemented. The system is designed to study the feasibility of the recognition of handprinted text in a loosely constrained environment. The NIST handprint database, NIST Special Database 1, is used to provide test data for the recognition system. The system consists of eight functional components. The loading of the image into the system and storing the recognition results from the system are I/O components. In between are components responsible for image processing and recognition. The first image processing component is responsible for image correction for scale and rotation, data field isolation, and character data location within each field; the second performs character segmentation; and the third does character normalization. Three recognition components are responsible for feature extraction and character reconstruction, neural network-based character recognition, and low-confidence classification rejection. The image processing to load and isolate 34 fields on a scientific workstation takes 900 seconds. The same processing takes only 11 seconds using a massively parallel array processor. The image processing components, including the time to load the image data, use 94 of the system time. The segmentation time is 15 ms/character and segmentation accuracy is 89 for handprinted digits and alphas. Character recognition accuracy for medium quality machine print is 99.8. On handprinted digits, the recognition accuracy is 96 and recognition speeds of 10,100 characters/second can be realized. The limiting factor in the recognition portion of the system is feature extraction, which occurs at 806 characters/second. Through the use of a massively parallel machine and neural recognition algorithms, significant improvements in both accuracy and speed have been achieved, making this technology effective as a replacement for key data entry in existing data capture systems.