Recognition experiments of cursive dynamic handwriting with self-organizing networks

摘要

The aim of the paper is to assess the feasibility of using self-organizing methods for the development of a recognition system of cursive, dynamic handwriting for interactive applications with a large dictionary. A prototype system based on segmentation into motoric strokes and concurrent segmentation/classification of allographs by means of a set of allographic maps is developed. In the initial implementation, based on Kohonen's SOMs (self-organized maps), a 70% user-specific word recognition rate with a 4k-words dictionary is approached. From this, indications are derived for a modified neural recognizer (SOC: self-organized classifier) that is still based on self-organization but is more flexible (dynamic network size and topology) and can support incremental learning. The new model, together with improved pre-processing methods, could overcome the 80% mark in a pilot study with three subjects.