Generic texture analysis applied to newspaper segmentation

This paper deals with the segmentation of grey-scale newspaper images into the distinct regions of text, picture and background. Feature vectors are obtained from the image by analysing localised textual characteristics and textual variation. Analysis is performed in a generic way making few contextual assumptions. The contrast, size, orientation and resolution of the image are accounted for by a combination of this feature extraction process and subsequent parametrisation. Given such implicit invariance we are able to perform an initial point based classification through the use of a modified quadratic neural network. The use of a simple flood fill based algorithm allows the successful segmentation of newspaper images into distinct rectangular regions. Results of newspaper segmentation show the effectiveness of these methods