Unmixing and pigment identification using visible and short-wavelength infrared: Reflectance vs logarithm reflectance hyperspaces

Abstract

Hyperspectral imaging has recently consolidated as a useful technique for pigment mapping and identification, although it is commonly supported by additional non-invasive analytical methods. Since it is relatively rare to find pure pigments in aged paintings, spectral unmixing can be helpful in facilitating pigment identification if suitable mixing models and endmember extraction procedures are chosen. In this study, a subtractive mixing model is assumed, and two approaches are compared for endmember extraction: one based on a linear mixture model, and the other, nonlinear and Deep-Learning based. Two spectral hyperspaces are used: the spectral reflectance (R hyperspace) and the -log(R) hyperspace, for which the subtractive model becomes additive. The performance of unmixing is evaluated by the similarity of the estimated reflectance to the measured data, and pigment identification accuracy. Two spectral ranges (400 to 1000 nm and 900 to 1700 nm) and two objects (a laboratory sample and an aged painting, both on copper) are tested. The main conclusion is that unmixing in the -log(R) hyperspace with a linear mixing model is better than for the non-linear model in R hyperspace, and that pigment identification is generally better in R hyperspace, improving by merging the results in both spectral ranges.