Handwritten Manchu Historical Document Segmentation with Anisotropic Gaussian Kernel

Abstract

It is key technology for handwritten historical document image analysis to segment text lines. There is no doubt that it is very difficult for text line segmentation owning to varied forms of handwritten Manchu historical document images. A novel method for handwritten Manchu historical document segmentation is presented. Initially, this method extracts the blob lines of the text lines by anisotropic Gaussian kernel and obtains connected components by contours. Then, it breaks the adhesion region between lines which is obtained by blob lines merging. Finally the detected text line and connection components combine to form the initial text line region, and text lines segmentation is achieved by boundary area growing. Results of segmentation for heterogeneous documents are provided, and the proposed method is compared with projection profit and seam craving methods in the same settings. The method is good at handling the skew and adhesion Manchu text lines. Tested on dataset of handwritten Manchu historical documents, the recall of 97.57% correct lines is achieved by the proposed method.