Can deep learning assist automatic identification of layered pigments from XRF data?

Abstract

X-ray fluorescence spectroscopy (XRF) plays an important role for elemental analysis in a wide range of scientific fields, especially in cultural heritage. XRF imaging, which uses a raster scan to...

Atomic spectrometry update: review of advances in the analysis of metals, chemicals and materials

“…The same research group proposed an end-to-end pigment identification framework, including pigment library creation, XRF spectra simulation, mock-up preparation, a pigment identification deep learning model and a 2D pigment map generation . 327 As a case study, the framework was applied to late 19th/early 20th-century paintings by Paul Gauguin and Paul Cezanne. The former painting had been analysed previously using a combination of XRF, reflectance imaging spectroscopy and cross-section analysis and so a reliable dataset already existed for comparison.…”

Neural Networks for Hyperspectral Imaging of Historical Paintings: A Practical Review

“…Apart from paint component mapping, NN implementation has also been explored to solve other spectral unmixing problems related to the layered structure of paintings, such as thickness estimation and paint segmentation. Xu et al have targeted the multi-layered structure of the painting, andapplied a NN to fully automate the pigment identification process [59]. Their work is based on XRF data and has simulated XRF spectra with threelayered pigments to train the model.…”

“…The spectral range of the input data varies from RIS to macro-XRF, and the data samples are characterised with improved spectral resolution down to the nanometre scale and increased wavelength channels up to hundreds or even thousands. Some of the studies are based on XRF data, acquired in various experimental conditions (power source, acquisition time, and beam size) [50,51,59]. The RIS data cover several wavelength domains, from the most conventional VIS domain (380-750 nm [62], 400-700 nm [64], 400-720 nm [49]), extended to NIR (383-893 nm [56,57], 377-1033 nm [63], 377-1037 nm [61], 400-950 nm [53], 400-1000 nm [52,55], 822-1719 nm [54]), to the recently emerging SWIR range (930-2500 nm [52], 1000-2500 nm [27,58]).…”

“…Jones et al applied this method to create a training dataset containing 3000 spectra of 15 pigment classes [50], modelled as a simplified case that only accounted for the primary fluorescence of single-layered samples. Xu et al [59] have employed a more complex model that simulates XRF spectra for three-layer structures in mixtures of 12 different pigment classes, resulting in 16,224 spectra in total. Both studies have adopted a transfer learning strategy that further fine-tunes the NN model with a smaller quantity of experimental data, after the initial training with the large synthetic dataset.…”

“…The second type of learning, unsupervised learning, is used to analyse and cluster unlabelled data [51,52,64]. In the case of transfer learning, the model is pre-trained with a large dataset and then fine-tuned with a separate smaller dataset [50,59,61]. This type of learning is discussed in detail in Section 5.3.2.…”