Segmentation Based Online Word Recognition: A Conditional Random Field Driven Beam Search Strategy

In this paper we undertake recognition of online unconstrained cursive handwritten English words. In contrast to popular dynamic programming or HMM-based approaches we propose a Conditional Random Field (CRF) driven beam search strategy applied in a combined segmentation-and-recognition framework. First, a candidate segmentation lattice is built using over-segmented primitives of the word patterns. Recognition is accomplished by synchronously matching lexicon words with nodes of the lattice. Probable search paths are evaluated by integrating character recognition scores with geometric and spatial characteristics of the handwritten segments in a CRF (conditional random field) model. To make computation efficient, we use beam search to prune the set of likely search paths. This overall system has been benchmarked on a new publicly available dataset - IBM_UB_1 as well as on the existing UNIPEN dataset for comparison.