Evaluation of pattern recognition techniques for the attribution of cultural heritage objects based on the qualitative XRF data

Abstract

Various instrumental and experimental modifications of portable XRF (pXRF) spectrometers meet nearly all the requirements for safe and efficient analysis of objects of cultural heritage, allowing this technique to be the most used analytical method for that purpose. The technique was used in a joint campaign (involving scientists, curators and art historians) to examine and protect an iconostasis containing well-preserved and reliably dated icons from four different centuries. The number of acquired XRF spectra was sufficient to be treated with pattern recognition techniques. Different pattern recognition techniques were applied in the analysis of cultural heritage object, both unsupervised and supervised. The results of various analytical examinations (non-invasive or performed on samples) pretreated or not, even a fusion of different techniques, were subjected to pattern recognition analysis.The possibilities of select pattern recognition techniques to analyze data obtained by a pXRF spectrometer from the surfaces of paintings and icons are presented in this study. Principal Component Analysis (PCA), Hierarchical Cluster Analysis (HCA), and Scattering Matrix Based Dimension Reduction (SMBDR) were used to process generated datasets of peak areas and spectral intensities providing dimension reduction of the initial datasets. The efficiency of this process is evaluated and discussed using the calculated index of informativeness and Bhattacharyya’s distance. Based on the results, the possibility of classification is discussed in terms of choosing proper classifiers, which were designed thereafter. The classification is evaluated by recognition ability and prediction ability parameters, and then tested by the leave-one-out procedure and using pXRF spectra collected from different color and support (cardboard paintings). An efficient procedure was developed for the attribution of artwork origin, based solely on XRF spectral data and PCA dimension reduction, followed by linear classification.