Adaptive algorithm for pigment identification from unmixing spectral data: Case study with two versions of a XVI century painting

Abstract

Artists commonly use a relatively reduced palette of pigments and mix them in different proportions to increase the gamut of colors present on artworks. In this study, a complete workflow for pigment identification using spectral unmixing of reflectance spectra in the visible and near infrared is presented. The algorithm includes superpixel segmentation as pre-processing to reduce the number of spectra that are unmixed. Then, a pre-extracted set of relevant color instances from the painting is used to build an adaptive subset of candidate pigments from a reference palette, and pigment identification is achieved by superpixel voting within the reduced subsets corresponding to the automatically extracted endmembers presence maps. Two different moments in time of a Maternity of the 16th century (original and restored) and a modern replica of the same painting are used to showcase the performance of the algorithm, which is able to correctly identify 80 % of the pigments present from a reference library of 23 pigments, taking less than three minutes for processing around 7000 spectra.