Recognition algorithm for handprinted Chinese characters by 2D-FFT

In this paper, we propose a new method for recognition of handprinted Chinese characters. The method devises to absorb the local variation in handprinting by using the property that the power spectral density is invariant under a translational transform. We compute a power spectral density at each local area (called segment) of a character image with the same size as the elementary picture, and obtain normalized correlation function between the power spectrum at each segment and that of elementary picture. We construct the feature vector by enumerating the normalized correlation functions with respect to segments and elementary pictures. The performance of the present method is superior to that of the directional pattern matching method which is a representative traditional method