Texture classification of photographic papers: improving spectral clustering using filterbanks on graphs

Abstract

From the point of view of graph signal processing, we show that spectral clustering is equivalent to an ideal low-pass filterbank. Building upon previous multiscale community detection ideas [11], and integrating the concept of community cores [8], we propose a data-driven filterbank-based classification method. We apply this method to the texture classification of photographic papers useful to art historians, and we show that it provides a richer and more informative description of the data’s structure in clusters.