Inkball models for character localization and out-of-vocabulary word spotting

Inkball models have previously been used for keyword spotting under the whole word query-by-image paradigm. This paper applies inkball methods to string-based queries for the first time, using synthetic models composed from individual characters. A hybrid system using both query-by-string for unknown words and query-by-example for known words outperforms either approach by itself on the George Washington and Parzival test sets. In addition, inkball character models offer an explanatory tool for understanding handwritten markings. In combination with a transcript they can help to to attribute each ink pixel of a word image to specific letters, resulting in high-quality character segmentations.