Data fusion and dual-domain classification analysis of pigments studied in works of art

Abstract

We propose a new mid-level data-fusion system to process, as a unique signal, the Raman and X-ray fluorescence (XRF) spectra obtained from the first micro-Raman–XRF instrument. The system is based on the main advantage of the wavelet transform, which is multiresolution. First, each spectrum set is split into blocks according to their frequency. The blocks which contains background and noise signals are removed and variable selection is performed on the remaining blocks to extract those variables with the most power of classification. These variables are concatenated and form a Raman–XRF meta-signal ensemble. Finally, dual-domain signal ensembles from references and samples are classified using partial least squares discriminant analysis (PLS-DA). Our results show that this system is suitable for rapidly and automatically classifying ancient pigments using the complementary information provided by both techniques. Classification with different levels of difficulty can be handled and no prior knowledge of the sample composition is required. This system has been applied to real spectra of ancient pigments and can also be applied to combinations of other spectral signals.