Combination or Segmentation-based and Wholistic Handwritten Word Recognition Algorithms

An alf&ithm for the recognition of unconstrained handwritten words is proposed. Based on an analysis of writing styles, it is shown that techniques for isolated character recognition. segmentation. as welJ as cursive script recognition are needed to achieve a robust solution to handwritten word recognition. A combination of these algorithms is proposed in which each method outputs a ranking of the words in the dictionary that are subsequently combined to generate a single consensus ranking. Preliminary results of the implementation of this methodology are given along with future research directions. 1. Introdudion An algorithm for handwritten word recognition must be able to successfully recognize the image of any word whether it is discretely printed. written cursively, or composed of a mixture of bOth styles. The writing styles that can be used to form a handwritten word are illustrated in Figure l. Discrete characters, cursive fragments (groups of characters written with a single continuous motion), and complete cursive words are often used either singly or in combination. TIie algorithmic approaCh discussoo in this paper is directed toward p~ial addresses. The handwritten words that occur in addresses are completely unconStrained by writer, style, instrument, size of text, placement within an image, and so on. However, one very irnporfaJrt constraint is that the words typically come from a fixed vocabulary. For example, the name of a city may be one of oVer 30,000 possibilities. Also, if some digits of the postal code can be recognized. they can help to considerably reduce the size of the lexicon. Sometimes it may be possible to limit the choices for a city name to two or three candidates. International Workshop on Frontiers in Handwriting Recognition, Chateau de Bonas, France, Sept. 23-27, 1991, 229-240.