F.J. Ferri et al. (Eds.): SSPR&SPR 2000, LNCS 1876, pp. 757-766, 2000. © Springer-Verlag Berlin Heidelberg 2000 Segmentation of Text and Graphics/Images Using the Gray-Level Histogram Fourier Transform M.A. Patricio 1 , D. Maravall 2 1 Centro de Cálculo 2 Departamento de Inteligencia Artificial Universidad Politécnica de Madrid dmaravall@fi.upm.es Abstract. One crucial issue in automatic document analysis is the discrimination between text and graphics/images. This paper presents a novel, robust method for the segmentation of text and graphics/images in digitized documents. This method is based on the representation of window-like portions of a document by means of their gray level histograms. Through empirical evidence it is shown that text and graphics/images regions have different gray level histograms. Unlike the usual approach for the characterization of histograms that is based on statistics parameters a novel approach is introduced. This approach works with the histogram Fourier transform since it possesses all the information contained in the histogram pattern. The next and logical step is to automatically select the most discriminant spectral components as far as the text and graphics/images segmentation goal is concerned. A fully automated procedure for the optimal selection of the discriminant features is also expounded. Finally, empirical results obtained for the text and graphics/images segmentation using a simple three-layer perceptron-like neural network are also discussed. Keywords: Feature extraction and selection; Image analysis; Applications: automatic document analysis. 1. Introduction – The Gray Level Histogram as a Discriminant Tool for Text and Graphics/Images Segmentation Document image analysis is an active research and development field [1] in which pattern recognition techniques are of the greatest interest. One critical issue in the automatic analysis of digitized documents is the separation of text and graphics/images. The text regions of the document are usually analyzed by means of well-known OCR techniques, whereas the graphics and images are just codified in order to obtain optimal storage and retrieval of such information. This communication describes a novel method for the segmentation of text and graphics/images. This method exploits the empirical evidence that regions of text and regions of graphics/images have very different gray level histograms. As an illustration, figure 1 shows two examples. Notice the application of a window on the original digitized document in order to compute the brightness histogram in small portions of the whole document. The practical issues concerning the window size and the scanning process