Abstract

The main objective of this thesis is to develop a system to automatically segment and label a variety of real-life documents written in different languages. The main idea is to partition the whole document into different subimages and assign to each of them one of two labels: text or non-text (including graphics); and then identify the text, as one of three categories, Roman; Ideographic, or Arabic script. The whole process consists of several steps. For instance, to detect the skew angle of the document, we use the Hough transform and the most frequently occurring local maximum. Moreover, in order to segment the page into regions, we have developed a novel approach based on diagonal scanning and node-edge orientation. Then, text and graphic components are also isolated using the geometric configuration of the connected components. Next, the textual components are segmented into lines using the projection profile: and finally the script is classified into one of the three categories mentioned above using the bounding boxes and horizontal projection. The system has been tested on 215 samples of diverse document types from many sources such as journal articles, magazines, newspapers, facsimiles, and office correspondence. These testing samples include low quality document images with different types of distortion; they also contain upside-down, skewed and low resolution images. The system classifies 93.5% of the script type correctly and 6.5% of these documents incorrectly