A variational bayesian style classification for typographic persian text using gabor features

The close visual relation between the style of typographic words in a document from one side and the conceptual meaning of `texture´ from the other side, has been used to propose an approach based on gabor filter extracted features to classify words in a document into three classes of regular, italic and bold. Since the generalized dirichlet distribution (GDD) is shown to be very flexible in image and texture modeling, we have chosen to fit a mixture of GDDs to the extracted feature space. Finally parameter estimation and classification is done using a variational bayesian method. Based on the obtained result, the performance of the proposed approach is demonstrated to be significant in classifying Persian words.