Word Level Discriminative Training for Handwritten Word Recognition

Word level training refers to the process of learning the parameters of a word recognition system based on word level criteria functions. Previously, researchers trained lexicon-driven handwritten word recognition systems at the character level individually. These systems generally use statistical or neural based character recognizers to produce character level confidence scores. In the case of neural networks, the objective functions used in training involve minimizing the difference between some desired outputs and the actual outputs of the network. Desired outputs are generally not directly tied to word recognition performance. In this paper, we describe methods to optimize the parameters of these networks using word level optimization criteria. Experimental results show that word level discriminative training without desired outputs not only outperforms character level training but also eliminates the difficulty of choosing desired outputs. The method can also be applied to all segmentation based handwritten word recognition systems.