Document Segmentation using Texture Segmentation

The alphabet in a script is characterized by a particular organization of line segments and curve segments. A script is distinguished by the nature of the organizational pattern for the primitives. These patterns occur at spatial frequencies which are typical to the script [16]. Text can thus be considered to possesses certain frequency and orientation information. Thus, the text can be segmented be treating it as a separate type of texture. An algorithm based on this property of text has been proposed [13] for extraction of text from address blocks in the scanned images of letters and documents. The main steps are as follows:

1. Filter the input image through a bank of n even-symmetric Gabor filters, to obtain n filtered images
2. Compute the feature image consisting of the ``local energy" estimates over windows of appropriate size around every pixel in each of the filtered images.
3. Cluster the feature vectors corresponding to each pixel using a squared error clustering algorithm to obtain a segmentation of the original input image into K clusters or segments.

This algorithm treats the document images as a multi-textured region in which the text on the envelope defines a specific texture and other non-text contents including blank regions correspond to different textures. Thus, the problem of segmenting a document is posed as a texture segmentation problem A similar approach can also be followed for the segmentation of multi-script documents as different scripts also have different responses to a bank of Gabor filters. For example, English and Malayalam have a high response to the 0 degree Gabor filter, because of the large number of straight vertical segments in their scripts. English also shows high response to the 90 degree Gabor filter because of horizontal line segments present in English. Devanagari scripts which have a headline, respond sharply to the 90 degree Gabor filter. Telugu on the other hand, which is mostly circular has nearly equal response in all the Gabor filter directions.