Texture feature characterization for logical pre-labeling

In this article we present a study based on the use of texture features for logical pre-labeling. The aim of our work is to calculate a great number of texture features over three sets of machine-printed document images and to study their joint discriminant power using SVM classifiers. The three corpuses we use are: the Archives of Savoie (AoS), composed of strongly structured documents, a subset of the UW3 database, and a third that is not structured at all, since it is composed of Web site images. The originality of our contribution is to sum up various methods that have been used for many years in our domain, and to test them on documents having very different specificities.