Character prototype selection for handwriting recognition in historical documents

Handwriting recognition in historical documents is vital for making scanned manuscript images amenable to searching and browsing in digital libraries. A valuable source of information is given by the basic character shapes that vary greatly for different manuscripts. Typically, character prototype images are extracted manually for bootstrapping a recognition system. This process, however, is time-consuming and the resulting prototypes may not cover all writing styles. In this paper, we propose an automatic character prototype selection method based on a forced alignment using Hidden Markov Models (HMM) and graph matching. Besides the predominant character shape given by the median or center graph, structurally different additional prototypes are retrieved with spanning and k-centers prototype selection. On the historical Parzival data set, it is demonstrated that the proposed automatic selection outperforms a manual selection for handwriting recognition with graph similarity features.