Optical Character Recognition Using Novel Feature Extraction & Neural Network Classification Techniques

This paper describes two novel techniques applied to the jkature extraction and pattern classification stages in an OCR system for typeset characters. A technique for estimating the class discrimination ability of continuous valued jkatures is presented leading to the formation of complex features which facilitate the classifimtion stage. Next, a neural network ClQSSafieT trained wing a nxently proposed powerfisl training algorithm, based m rigorous nonlinear programming methods, kz applied to large-scale OCR problems involving typeset Greek characters and found to exhibit good generalization capabilities compared to other conventional and artificial neural network (ANN) classifiers. Combining these jkature extraction and classification techniques in a unified software platform, we have designed an OCR system which achieved high mognition rates in some real world OCR ezperiments.