Segmentation of Tri-Lingual Documents

Abstract

Abstract-Physical layout analysis intends to study the arrangement of layouts or locations of the regions present in a document image before understanding it. Before extracting the text or information from a document image, page segmentation (layout analysis) techniques need to be applied to identify the exact layout (area) where the text or image resides. In Page Segmentation, Top-down methods are simple and efficient but fail in non Manhattan layouts. In contrast, Bottom-up approaches adapt non Manhattan layouts easily than the top down approaches, but heavily depend on the threshold, parameters and extensive computations for layout identification. On the other hand, Hybrid methods (Bruel [31], Bruel [32]) suits well for layout identification by eliminating the dependency on threshold and parameters. But this analyzes the white background of the image with small white rectangles and merges them to locate the content blocks. Merging of small white rectangles makes the identification process tedious since large number of small white rectangles gets involved in the image. In addition, this approach heavily relies on heuristics for merging operations, which affects the segmentation rate considerably. In all the above reported methods (Bottom up and Hybrid approaches), connected component analysis (requires more number of pixel visits) is required to identify black and white components from the image. Therefore, the above shortcomings motivated this research towards designing a White Space analysis technique which eliminates the usage of the connected component analysis (to identify white spaces), heuristics, threshold and prior knowledge. As a result, in this thesis, Rectangular White Space Analysis (RWSA) technique has been proposed to grab all the white spaces over the image in a single scan over the image with minimum pixel visits, and the white spaces are merged together without the assumptions of heuristics and threshold to segment the layouts. Moreover, two statistical properties have also been proposed in this thesis, to separate the text blocks and images from the identified layouts and this hybrid approach has been explained in the subsequent section.