Discriminative Bernoulli Mixture Models for Handwritten Digit Recognition
Hidden Markov Models (HMMs) are now widely used in off-line handwriting word recognition. In contrast to the conventional approach, based on Gaussian mixture HMMs, we have recently proposed to directly feed columns of raw, binary pixels into Bernoulli mixture HMMs, which are extracted using a sliding window of adequate width. In this work we propose two improvements to this approach. On the one hand, we propose to repositioning in each column the sliding-window according to its center of mass. On the other hand, we propose to transform a BHMM classifier into an equivalent log-linear model with hidden variables in order to performe a discriminative training.