Simon BRENNER | Bernadette FRÜHMANN | Willi VETTER | Federica CAPPA | Manfred SCHREINER | Heinz MIKLAS | Robert SABLATNIG
(TU Wien, Vienna, Austria)

Keywords: multispectral imaging spectroscopy database

The research groups INTK (Academy of Fine Arts Vienna) and CVL (TU Wien) have a long history of collaborative work on historic manuscript analysis. While CVL is concerned with Multispectral Imaging (MSI) and document analysis, INTK performs non-invasive material analysis by combining complementary spectroscopy methods.
In the course of four joint projects conducted during the last ten years, around 100 manuscripts were examined. For half of them, both multispectral imagery and spectroscopic data were collected. Until now, the analysis of the acquired data was performed in large part independently by the respective groups. Now we are at a point where enough data for systematic analysis of the previously examined manuscripts is possible. This, however, requires a thoroughly structured database.
We will present our attempts to create a unified database containing imagery, measured spectra and detected pigments as well as metadata about both the investigated object itself and the technical methods that have been applied. Such a database will support the following applications:
Retrieval. Documents can be queried for similar properties derived from MSI or spectroscopy. This allows to study manuscripts on a meta-level and find possible connections between different manuscripts.
Machine learning. Finding correlations between MSI and spectroscopic measurements could help to establish rules for the discrimination of certain similar pigments from multispectral data alone. Furthermore, the application of machine learning to the determination of pigments from spectroscopic measurements, which is currently performed by experienced experts, will be evaluated.
Interactive visualization. Point measurements can be visualized on top of multispectral images, along with spectral curves and detected materials, providing a more intuitive interface to the data. Further, using correlations between MSI and spectroscopy as mentioned in the previous paragraph, point measurements can be extrapolated across larger domains, thus producing approximate element maps without the need to perform a spectroscopic scan.