Layout analysis of book pages

A method is proposed for analysing the geometric and logical structure of pages in a typical single-column book. A Gaussian blur combined with thresholding is used to form connected components which nominally represent words. A bottom-up nearest-neighbour approach is used to find textual lines, and a manually-defined line length parameter is used to remove marginal noise and find the page frame. A state machine is used to group lines and label them according to function. The proposed method is able to correctly segment and label 99.82% of all targeted features in a set of 196 sample pages. Of the sixteen errors encountered in the sample pages, eleven are instances where adjacent lines have been merged together, four are instances where paragraphs have been split in half, and the remaining error was caused by a header element being detected as part of the body text.