Combining High-Level Features with Sequential Local Features for On-Line Handwriting Recognition

The trade-oo between high-level, long-range features and low level, local features is common among many pattern recognition problems: the former are usually more powerful but less robust, while the latter is less informative but more reliable. In this paper we describe a new method for combining high-level long-range features and local features for on-line handwriting recognition. First, high-level features such as crossings, loops and cusps are extracted. A localization procedure is then applied to spread these high-level features over the neighboring sample points, resulting in local representations of nearby high-level features. These features are then combined with the usual local features at each sample point and used in an integrated segmentation and recognition process. This method allows incorporation of information carried by high-level long-range features while at the same time maintains the high reliability of the recognition system. We report experimental results on an HMM based recognizer for writer independent recognition of unconstrained handwritten words.