Layout Analysis and Text Column Segmentation for Historical Vietnamese Steles

Stone engravings in Historical Vietnamese steles allow historians to study the life of common people in the villages. Only recently, a large amount of images of such engravings have become available. For supporting the historians, automatic document analysis systems are needed for reading the ancient Chu Nôm characters that are written in columns from top to bottom. In this paper, we study the problem of layout analysis, which is the first step of automatic reading. Semantic segmentation is applied at pixel-level to find the title, main text, label, and reference number on the page using deep convolutional neural networks. Afterwards, seam carving is used to segment the text columns within the main text. We present baseline results for hundred exemplary pages, discuss error cases, and outline lines of future research.