Connected pattern segmentation and title grouping in newspaper images

Abstract

This paper presents an algorithm that performs automated segmentation and classification of newspaper images. The algorithm discusses a technique for segmenting components that are connected to other components and presents another technique to correctly group titles and subtitles. The algorithm uses a bottom-up approach to initially segment the image, classify patterns and extract text lines. The classified patterns are then merged into complete regions. The algorithm is tested on a set of complex newspaper images taken from the First International Newspaper Segmentation Contest, and the results are compared with the contest results.