Classification of elementary stamp shapes by means of reduced point distance histogram representation

Abstract

The paper presents a problem of stamp shape classification, where an input stamp is given as a bitmap containing binary values. While every stamp features a specific geometrical form coming from the de facto standards of stamping process, thus it can be classified as round, oval, square, rectangular or triangular. We assume to have a detected stamp and in this paper we deal with the stage of features extraction and reduction, by means of Point Distance Histogram (at the stage of features extraction) and Principal Component Analysis and Linear Discriminant Analysis (at the stage of dimensionality reduction). The final classification employs similarity evaluation involving hand-drawn templates, ideal shapes and average descriptors calculated for the entire database. Despite the fact that there are only several basic stamp shapes, the task is not trivial since there are many variations in size, silhouette and complexity of individual stamps. It should be emphasized that the scanned document may be degraded in quality thus extracted stamp can be distorted (the silhouette may be discontinuous and/or can be noised). The paper provides some experimental results on real documents with different types of stamps and a comparison with a classical Discrete Cosine Transform (DCT) and PCA applied on image matrix. © 2012 Springer-Verlag.