Structures of Covariance Matrix in Handwritten Character Recognition

The integrated approach is a classifier established on statistical estimator and artificial neural network. This consists of preliminary data whitening transformation which provides good starting weight vector, and fast training of single layer perceptron (SLP). If sample size is extremely small in comparison with dimensionality, this approach could be ineffective. In the present paper, we consider joint utilization of structures and conventional regularization techniques of sample covariance matrices in order to improve recognition performance in very difficult case where dimensionality and sample size do not differ essentially. The techniques considered reduce a number of parameters estimated from training set. We applied our methodology to handwritten Japanese character recognition and found that combination of the integrated approach, conventional regularization and various structurization methods of covariance matrix outperform other methods including optimized Regularized Discriminant Analysis (RDA).