A method for acquiring and modeling color-document images

We propose a method of pre-processing of document images when a classical grey-levels acquisition is not sufficient, in the aim to extract text blocks, or to understand the physical structure of the document. From an RVB acquisition of the document, we search, by apprenticeship, to obtain the most informative color features for the segmentation of a given kind of document. The processing is performed by a background filtering and an extraction of the informative areas, thanks to a 2D-AR modeling. The segmentation is made by the use of dissimilarity measures between the laws followed by the prediction errors. In this study, we use by three times an Akaike criterion: a first time to find which are the most informative color features, then to determine the optimal order of the AR model, then finally, to approximate the laws of the prediction errors by optimal histograms, allowing a data reduction