Recognition-based system for segmentation of handwritten numeral strings
Yun Lei · Journal of Tsinghua University(Science and Technology) · 2005
A recognition-based system was developed for segmentation of handwritten numeral strings to deal with the recognition problems when the numerals touch each other. External contour analysis and projection analysis are combined to locate the candidate segmentation lines. These candidate segmentation lines are then used to over-segment the numeral string. Each subimage 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. Then, all the candidate segmentation results are described in a probabilistic model with a single-digit classifier used to recognize each clique. Finally, maximum a posterior (MAP) criterion is used to select the optimal segmentation result from all candidate segmentation results. The search for the optimal result uses a pruning algorithm to reduce the time and space complexity for real-time applications. Test results on collections of samples from NIST SD19 show that the system achieve an accuracy of 97.72%.