Character recognition using canonical invariants

Abstract

This paper presents a new insight into character recognition problem. Implicit polynomial (IP) curves have been used for modelling characters. A unique decomposition theorem is employed to decompose these curves into simple line primitives. For the comparison of the characters, canonical invariants have been computed using so called related points of the curves, which are the real intersections of the lines. Experimental results are presented to asses discrimination power of proposed invariants and their robustness under data perturbations. The method has also been compared with fourier descriptors.