Efficient Stamps Classification by Means of Point Distance Histogram and Discrete Cosine Transform

The problem of stamp recognition addressed here involves a multi-stage approach which includes stamp detection, localization and segmentation, features extraction and finally, classification. In this paper we focus on the two last stages, namely features extraction by means of Point Distance Histogram and Discrete Cosine Transform, and classification employing distance calculation by means of Euclidean metrics. The first stage which leads to automatic stamps segmentation has been described in several previous papers and it is based mainly on color segmentation. The feature extractor described here works on binary images of stamps and employs polar representation of points gathered in a histogram form, which is later reduced by means of Discrete Cosine Transform. At the classification stage, compact descriptors of stamps are compared according to the distance to the reference objects (class’ centers), and the closest class is taken as the answer. The paper includes some results of selected experiments on real documents having different types of stamps. A comparison with the classical two-dimensional DCT calculated over the images is also provided to prove high discriminative power of the developed approach.