Page segmentation for document image analysis using a neural network

In this paper we present a method for segmenting document page images into text and nontext regions. The underlying assumption made by this approach is that the two regions can be viewed as different textures. We do not use any a priori knowledge of the document format. A convolution-based method is used to generate the texture feature images. The coefficients of the convolution masks are obtained using a single-layer artificial neural network that generates eigenvectors of the correlation matrix of the input data. The coefficients of these masks have been ''learned'' from examples of the document images and have a potential of being considerably more powerful than masks with preset coefficients. A thresholding scheme based on a measure of entropy is used to Segment the feature images into the homogeneous regions. (C) 1996 Society of Photo-Optical instrumentation Engineers.