Approximative Bayes optimality linear discriminant analysis for Chinese handwriting character recognition

Abstract

Discriminant subspace learning is an important branch for pattern recognition and machine learning. Among the various methods, Bayes optimality linear discriminant analysis (BLDA) has shown its superiority both in theory and application. However, due to the computational complexity, BLDA has not been applied to large category pattern tasks yet. In this paper, we propose an approximative Bayes optimality linear discriminant analysis (aBLDA) method for Chinese handwriting character recognition, which is a typical large category task. In the aBLDA, we first select a set of convex polyhedrons that are obtained by the state-of-the-art methods, then the searching zones are limited to these polyhedrons. Finally, the best of them is chosen as the final projection. In this way, the computational complexity of BLDA is reduced greatly with comparable accuracy. To find more than 1D projections, the orthogonal constraint is employed in the proposed method. The experimental results on synthetic data and CASIA-HWDB1.1 show the effectiveness of the proposed method.