An analytical handwritten word recognition system with word-level discriminant training

We describe an analytical handwritten word recognition system combining neural networks (NN) and hidden Markov models (HMM). Using a fast left-right slicing method, we generate a segmentation graph that describes all possible ways to segment a word into characters. The NN computes the observation probabilities for each character hypothesis in the segmentation graph. Then, using concatenated character HMMs, a likelihood is computed for each word in the lexicon by multiplying the observation probabilities over the best path through the graph. The role of the NN is to recognize characters and to reject non-characters. We present our approach to globally train the word recognizer using isolated word images. Using a maximum mutual information (MMI) cost function at the word level, the discriminant training updates the parameters of the NN within a global optimization process based on gradient descent. The recognizer is bootstrapped from a baseline recognition system, which is based on character level training. The recognition performance of the globally trained system is compared to the baseline system