A two-stage approach for segmentation and recognition of handwritten digit strings collected from mail pieces

We present an approach to interpret handwritten digit strings by employing a two-stage segmentation and recognition scheme. The first stage processing is to deal with possible touching between digits. An input string is represented with an abstractive means, called primitive. Candidate segmentation points are sought from the representation. To locate a boundary between digits, a recognition engine, which is rough but fast, is applied to a combined segment. The second stage recognition process, which is more accurate but slow, is applied to the combined segment depending on the recognition confidence of the first stage process. Features for the segmentation and recognition are defined and described. To evaluate the effectiveness of the approach, experiments have been performed on a set of digit string images collected from real mail pieces, and promising preliminary results have been observed.