Segmentation of Overlapping Digits through the Emulation of a Hypothetical Ball and Physical Forces

Abstract

This paper presents an algorithm for segmenting pairs of overlapping handwritten digits. Digits can be found overlapped in text depending on writing style and organization; digits in close proximity or with elongated strokes may also overlap with their neighbors. Applications such as automated character recognition are directly affected by overlapping characters and their segmentation. The proposed approach is based on the emulation of inertia and a deformable hypothetical ball. The strokes act as a pathway for the ball to run and create the segmentation. The results of the algorithm are subject to a digit recognizer and it is shown that the method performs well and presents lower computational cost when compared to other segmentation approaches.