Statistical Feature Fusion for Sassanian Coin Classification

Abstract

Ancient coins classification has attracted increasing attention for the benefits which it brings to numismatic community. However, high between-class similarity and, in the meantime, high within-class variability make the problem a particular challenge. This issue highlights the importance of extracting discriminative features for ancient coins classification. Therefore, in this paper, the capability of statistical feature fusion was examined. First, a representation of the coin image based on the phase of the 2-D Fourier transform of the image is using so that the adverse effect of illumination was eliminated. The phase of the Fourier transform preserves the locations of the edges of a given coin image. The problem of unwrapping is avoided by considering two functions of the phase spectrum rather than the phase directly. Then, BDPCA approach which can reduce the dimension of the phase spectrum in both column and row directions is used and an entry-wise matrix norm calculates the distance between two feature matrices so as to classify coins. Extensive experiments are conducted on a database of Sassanian coins in order to compare the performance of proposed method with the other feature extraction method which are used in other works. The results show the proposed method is promising.