Line segmentation for grayscale text images of khmer palm leaf manuscripts

Abstract

Text line segmentation is one of the most essential pre-processing steps in character recognition and document analysis. In ancient documents, a variety of deformations caused by aging produce noises which make the binarization process very challenging. Moreover, due to the irregular layout such as skewness and fluctuation of text lines, segmenting an ancient manuscript page into lines still remains an open problem to solve. In this paper, we propose a novel line segmentation scheme for grayscale images of Khmer ancient documents. First, a stroke width transform is applied to extract connected components from the document page. The number and medial positions of text lines are estimated using a modified piece-wise projection profile technique. Those positions are then modified adaptively according to the curvature of the actual text lines. Finally, a path finding approach is used to separate touching components and also to mark the boundary of the text lines. Experiments are conducted on a dataset of 110 pages of Khmer palm leaf manuscript images by comparing the robustness of the proposed approach with existing methods from the literature.