1-dimensional and pseudo 2-dimensional HMMs for the recognition of German literal amounts

Hidden Markov models (HMMs) are frequently used in off-line cursive script recognition. In most cases, the script is processed strictly from left to right, yielding a sequence of feature vectors fed into the HMM recognizer. In order to achieve good recognition results, more or less complex normalization has to be performed on the script beforehand to reduce the effects of writer variability. Taking the example of German literal amounts, ruler-line estimation and height normalization in particular become difficult tasks, due to the often extremely long words. In many cases, the assumption of straight ruler lines does not hold. Thus, it would be advantageous to integrate normalization into the recognition process. We present two approaches using regular 1D HMMs in comparison to pseudo-2D HMMs (P2DHMM). Pre-processing is basically identical for both approaches. In the case of P2DHMMs, however, we use a much simpler ruler line estimation scheme. Under otherwise similar conditions, the P2DHMMs have proved to be slightly superior to the regular HMMs in initial experiments.