Recognition of connected numeral strings using partial boundary features

Connected or touching numerals appear very often in handwritten documents when they are to be processed by a recognition system. They could be the results of writing habit, limited resolution of the graphic input device, or some preprocessing of the image. In this paper the author describes an approach to identifying connected numerals without a priori knowledge of the number of the numerals in the string. The left boundary of the numeral cage is used for the classification. Special methods are developed for determining the left boundary and extracting the features on it.<>