A connectionist system for learning and recognition of structures: application to handwritten characters

Abstract

A connectionist system for learning and recognition of structures is developed. The system is a cascade of two different modules, one for detecting linear structures (primitives) and the other for integrating these linear structures. A connectionist model implementing Hough transform has been used for the first module. The peaks in the Hough space are found by iterative verification method. A multilayered perceptron (four layers) with suitably chosen number of nodes and links has been used for the second module. As long as the size of the output layer of first module remains fixed (even if the size of input image changes), the same second module can be used and this is because the modules operate independently. The system performance is tested on handwritten Bengali character set.