Adaptive normalization of handwritten characters using GAT correlation and mixture models

This paper proposes an adaptive or category-dependent normalization technique for handwritten characters featuring global affine transformation (GAT) correlation and mixture models. Key ideas are twofold. First, we estimate a probability density function (PDF) of black pixels for each category using mixture models of Gaussian distribution functions and the EM algorithm. Second, we determine optimal, global affine transformation that maximizes a normalized cross-correlation value between a GAT-superimposed input pattern and the above-mentioned PDF by the successive iteration method. Experiments using the handwritten numeral database IPTP CDROM1B show that the entropy of optimally GAT-superimposed test samples decreases substantially by more than 20%. We discuss the enhanced normalization ability and the computational complexity of the proposed method.