Segmentation of the Date in Entries of Historical Church Registers

Handwriting recognition requires a prior segmentation of text lines which is a challenging task, especially for historical scripts. Exemplary for the date in entries of historical church registers, we present an approach which enables a segmentation by using additional knowledge about the word sequence. The algorithm is based on probability distribution curves and a neural network, which assesses local features of potential word boundaries. Our database consists of 298 different date entries from the 18th and 19th century which contain 674 word boundaries. The algorithm generates hypotheses for the expected date type, ordered by their probability. Tests resulted in an accuracy of 97% for the best four hypotheses.