Neural networks with enhanced outlier rejection ability for off-line handwritten word recognition

Abstract For a segmentation and dynamic programming-based handwritten word recognition system, outlier rejection at the character level can improve word recognition performance because it reduces the chances that erroneous combinations of segments result in high word confidence values. We studied the multilayer perceptron (MLP) and a variant of radial basis function network (RBF) with the goal to use them as character level classifiers that have enhanced outlier rejection ability. The variant of the RBF uses principal component analysis (PCA) on the clusters defined by the nodes in the hidden layer. It was also trained with and without a regularization term that was aimed at minimizing the variances of the nodes in the hidden layer. Our experiments on handwritten word recognition showed: (1) In the case of MLPs, using more hidden nodes than that required for classification and including outliers in the training data can improve outlier rejection performance; (2) in the case of PCA–RBFs, training with the regularization term and no outlier can achieve performance very close to training with outliers. These results are both interesting. Result (1) is of interest because it is well known that minimizing the number of parameters, and therefore keeping the number of hidden units low, should increase the generalization capability. On the other hand, using more hidden units increases the chances of creating closed decision regions, as predicted by the theory in Gori and Scarselli (IEEE Trans. PAMI 20 (11) (1998) 1121). Result (2) is a strong statement in support of the use of regularization terms for the training of RBF-type neural networks in problems such as handwriting recognition for which outlier rejection is important. Additional tests on combining MLPs and PCA-RBF networks showed the potential to improve word recognition performance by exploiting the complementarity of these two kinds of neural networks.