Skip to Main Content
A novel recognition-based system for segmentation of touching handwritten numeral strings is proposed. In this paper, we combine external contour analysis and projection analysis to find candidate segmentation points. With internal contour analysis, the candidate segmentation points is utilized to determine the corresponding candidate segmentation lines with which the numeral string is over-segmented. Each sub-image of the over segmented string is defined as a fragment. The combination of one or more adjacent fragments is defined as a clique. Thus, each candidate segmentation result is composed of one or more cliques. Subsequently, all the candidate segmentation results are described in a probabilistic model, and a classifier is embedded to recognize each clique. Finally, with the maximum a posterior (MAP) criterion, the optimal segmentation result is selected from all candidate segmentation results. This scheme is effective and robust for both single and multiple touching numerals. Experiment results on collection of samples from NIST SD19 show that our system can achieve a correct rate of 97.72% without rejection, which compares favorably with those reported in the literature.