Voting Based Text Line Segmentation in Handwritten Document Images

Abstract

Text line segmentation is a critical task in unconstrained handwritten document recognition. In this paper, a novel text line segmentation based on 2D tensor voting is proposed. 2D tensor voting is originally used to remove outliers and extract perceptual structures such as curves, junctions and end points from a set of sparse data points. Since characters of a text line are aligned on a smooth curve, 2D tensor voting is a useful tool for text line segmentation. First, center points of connected components generated from text pixels are encoded by second order tensors. These tensors then communicate with each other by a 2D stick voting process. Finally, the curve saliency values and normal vectors of resulting tensors are used to segment text lines. The experimental results obtained from ICDAR testing dataset show the effectiveness of our method.