Application of extended generalized linear discriminant functions (EGLDF) to optical character recognition

An extension to generalized linear discriminant functions, known as “EGLDF”, is applied to obtain very accurate classifiers for planar shapes based on contour coding and string edit distances. In this approach edit weights can be made dependent on the (local) “positions” of the prototypes to be matched with the test strings, thus allowing for very fine discrimination based on both global and local features of the shapes considered. Furthermore, the EGLDF framework provides effective techniques to optimally learn the required discriminative weights from training data, based on simple extensions of well-known gradient descent techniques such as the perceptron-pocket algorithm. The capabilities of the proposed approaches are assessed through classification experiments where the planar shapes correspond to images of handwritten digits from several writers which are represented by the chain codes of their contours