Unconstrained Handwritten Word Recognition Using a Combination of Neural Networks

Automatic offline recognition of text handwritten by different writers is still an open problem, due to several challenges including strong variability in writing styles, noise embedded in the environment, segmentation issues and others. In order to avoid errors during character segmentations, systems based on recognition of whole words have been developed lately. In this paper we present a novel method for classification of isolated handwritten words based on three components: a self organizing map (SOM) for non-supervised classification of segments of a word, a function measuring probabilities of each segment belonging to a specific cluster and a simple recurrent network (SRN) for temporal classification of a sequence of feature vectors obtained from segments forming the world. The experiments showed that the combination of these three components significantly improved the classification of words obtained from the benchmark IAM when compared with a multi-layer perceptron (MLP) and a plain combination of SOM and MLP. The proposed classifier obtained a mean word accuracy of 78.2% over a test set, compared to 66.2% obtained by a SOM combined to a MLP and to 32.1% obtained by MLP.