Segmentation of closely set and touching lines in handwritten document images using fringe maps

Abstract

Segmentation of handwritten text line is a challenge due to the large variety of situations that arise. The concept of a text line itself is difficult since a simple geometric concept like a bounding box for a line is not always applicable. Further, line skew may vary across a line, and also between lines. The problem in handwritten texts is also difficult since there is a large variation in spacing between lines in the writing. This paper presents a new method to address the issues of handwritten line segmentation. We propose the use of regions of white space amidst the text to bring out segmentation. The segmenting paths are found using the concept of fringe maps which are computed on the input binary image. Essentially the fringe value at any white space location helps to guide segmentation since it gives an indication of distance to text. Here we improve upon certain concepts of segmentation based on fringes and show several situations where this concept is successful and more natural than other approaches in the literature. Performance is compared with other approaches in literature on easily accessible ICDAR 2009 dataset.