Next Generation Optical Character Recognition using the Polynomial Method

This project extends on the work done Drs. Anderson and Gaborski’ in Optical Character Recognition (OCR) of handwritten digits done back in the 1990s. Their work involved using the polynomial discriminant method, iterative training, and genetic algorithms to optimize the feature selection. This project involves several variations in technique over previous work. The first is the use of the matrix manipulation language Matlab rather than ’C’ This allowed the faster training of the matrices (due to highly optimized matrix manipulation functions) and rapid debugging due to a built in debugger and other high level language advantages. This project also utilized a less random form of feature selection than the ”knight moves” or other feature selection techniques used previously. This technique involves compressing the digits from 30X20 pixels into 8X5 and 6X4 ”fat pixels” and then taking all possible combinations of pixels between the two compressed digits (resulting in 960 features) This project also explored the creation of a network of classifiers trained to differentiate between between all possible pairs of digits (0,1), (7,8), et. al. This resulted in 10C2 = 45 classifiers used in tandem to recognize the hand written digits. The Polynomial Classifiers initially created were very similar to those produced previously with one major exception: When the 960 features were reduced to 300 features using the genetic algorithm, the weight matrix proved to be surprisingly skilled at identifying previously unseen digits. For example, when classifying digits not used to train the matrix, the 960 features produced a 43% accuracy while the 300 features produced a 75% accuracy. By using a larger training set, and training for more iterations, this accuracy was improved to nearly 80% correctly classified digits. The 10C2 = 45 network classifier performed best of all classifying up to 83% of unseen digits correctly, opening up a whole new area for further research.