Short-Wave Infrared imaging to characterize historical Paintings – SWIP

Short-wave infrared range (SWIR, “extended NIR”, ca. 1000 to 2500 nm i.e. 10000 to 4000 cm-1) has a great potential for the characterization of art paintings, and in particular for the non-invasive identification of binders and organic materials, which is crucial for the authentication of works of art. However, this potential seems to be under-exploited, in particular due to the complexity of data interpretation. The aim of this project is to develop the use of SWIR spectroscopy and hyperspectral imaging for the study of historical paintings, taking into account their complex hybrid composition (superposition of layers, each prepared by grinding various mineral pigments and organic binders). To do so, I propose advanced multimodal acquisition combined to developments in data interpretation (thanks to the set-up of an extensive database) and data treatment procedures according to the specificities of the artwork and of this wavelength domain. This project will allow the implementation of a non-invasive methodology for the characterization of binders in historical paintings.

The first objective will be to characterize paint materials in the SWIR domain. ). A large set of references (pure pigments and paint, single and multi-layers) will be investigated to provide high-resolution SWIR spectra of the main components of historical paintings, and set an extensive database, not available in the community up to now. This will allow to investigate the absorption of common paint materials in the SWIR domain, and better underline inherent limitations in the characterization of pigments, binders and varnishes, particularly focusing on the identification of organic components.
We will then fully characterize and optimize our instrument capabilities. We propose an innovative methodology for spectral super-resolution, using dictionary learning to transfer the resolution of FORS system to the hyperspectral image.
Finally attention will be devoted to data treatment, which has to be adapted to the specificities of this wavelength range, as well as to the peculiar complexity of the systems investigated. Direct interpretation of SWIR spectra provides only limited information due to the overlap of the numerous overtones and combination bands. Detailed exploration of these spectra thus requires the combination of classical techniques such as the use of difference or derivative spectra, and /or more advanced tools including chemometric methods, spectral deconvolution, etc... Then, considering the huge amount of spectra created in few minutes by the SWIR hyperspectral camera, advanced statistical data treatment will be carried out on the data cube to extract both major and minor information, considering specific intensities variations. Common multivariate analyses (e.g., PCA, NNMA) will be combined and compared with innovative machine learning techniques (e.g. t-SNE, U-MAP), for the first time in the SWIR domain, in the case of subtle intensities variations.