Character prototyping in document images using Gabor filters

In this article we present a particular application of Gabor filtering for machine-printed document image understanding. To do so, we assume that the text can be seen as texture, characters being the smallest texture elements, and we verify this hypothesis by a series of experiments over different sets of character images. We first apply a bank of 24 Gabor filters (4 frequencies and 6 orientations) on each set, then we extract texture features further used to classify character images without a priori knowledge using a Bayesian classifier. Results are shown for different characters written in a same font, and for different font types given a character.