On the influence of vocabulary size and language models in unconstrained handwritten text recognition

In this paper we present a system for unconstrained handwritten text recognition. The system consists of three components: preprocessing, feature extraction and recognition. In the preprocessing phase, a page of handwritten text is divided into its lines and the writing is normalized by means of skew and slant correction, positioning and scaling. From a normalized text line image, features are extracted using a sliding window technique. From each position of the window nine geometrical features are computed. The core of the system, the recognizes is based on hidden Markov models. For each individual character, a model is provided. The character models are concatenated to words using a vocabulary. Moreover, the word models are concatenated to models that represent full lines of text. Thus the difficult problem of segmenting a line of text into its individual words can be overcome. To enhance the recognition capabilities of the system, a statistical language model is integrated into the hidden Markov model framework. To preselect useful language models and compare them, perplexity is used. Both perplexity as originally proposed and normalized perplexity are considered.