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 acquire spectra pixel-wise across artworks, provides the opportunity for spatial analysis of pigment distributions based on their elemental composition. However, conventional XRF-based pigment identification relies on time-consuming elemental mapping facilitated by the interpretation of measured spectra by experts. To reduce the reliance on manual work, recent studies have applied machine learning techniques to cluster similar XRF spectra in data analysis and to identify the most likely pigments. Nevertheless, it is still challenging to implement automatic pigment identification strategies to directly tackle the complex structure of real paintings, e.g. pigment mixtures and layered pigments. In addition, pigment identification based on XRF on a pixel-by-pixel basis remains an obstacle due to the high noise level. Therefore, we developed a deep-learning based pigment identification framework to fully automate the process. In particular, this method offers high sensitivity to the underlying pigments and to the pigments present in low concentrations, therefore enabling robust mapping of pigments based on single-pixel XRF spectra. As case studies, we applied our framework to lab-prepared mock-up paintings and two 19th-century paintings: Paul Gauguin's Poèmes Barbares (1896) that contains layered pigments with an underlying painting, and Paul Cezanne's The Bathers (1899-1904). The pigment identification results demonstrated that our model achieved comparable results to the analysis by elemental mapping, suggesting the generalizability and stability of our model.

X-ray fluorescence (XRF) spectroscopy is a common technique in the field of heritage science. However, data processing and data interpretation remain a challenge as they are time consuming and often require a priori knowledge of the composition of the materials present in the analyzed objects. For this reason, we developed an open-source, unsupervised dictionary learning algorithm reducing the complexity of large datasets containing 10s of thousands of spectra and identifying patterns. The algorithm runs in Julia, a programming language that allows for faster data processing compared to Python and R. This approach quickly reduces the number of variables and creates correlated elemental maps, characteristic for pigments containing various elements or for pigment mixtures. This alternative approach creates an overcomplete dictionary which is learned from the input data itself, therefore reducing the a priori user knowledge. The feasibility of this method was first confirmed by applying it to a mock-up board containing various known pigment mixtures. The algorithm was then applied to a macro XRF (MA-XRF) data set obtained on an 18th century Mexican painting, and positively identified smalt (pigment characterized by the co-occurrence of cobalt, arsenic, bismuth, nickel, and potassium), mixtures of vermilion and lead white, and two complex conservation materials/interventions. Moreover, the algorithm identified correlated elements that were not identified using the traditional elemental maps approach without image processing. This approach proved very useful as it yielded the same conclusions as the traditional elemental maps approach followed by elemental maps comparison but with a much faster data processing time. Furthermore, no image processing or user manipulation was required to understand elemental correlation. This open-source, open-access, and thus freely available code running in a platform allowing faster processing and larger data sets represents a useful resource to understand better the pigments and mixtures used in historical paintings and their possible various conservation campaigns.