1. Towards automatic classification of diffuse reflectance image cubes from paintings collected with hyperspectral cameras.
- Author
-
Kleynhans, Tania, Messinger, David W., and Delaney, John K.
- Subjects
- *
ART materials , *AUTOMATIC classification , *SPECTRAL reflectance , *PAINTING collecting , *CONVEX geometry , *CUBES , *MATCHING theory - Abstract
• While much progress has been made in developing hyperspectral cameras for collecting high quality reflectance image cubes from works of art, little work has been done to create optimized image analysis tools to mine these rich and large data sets. • Automated convex geometry algorithms are good candidates for pigment classification of reflectance imaging spectroscopy (RIS) data from paintings. • The proposed Maximum Distance (MaxD) algorithm performs well with automatic classification of paintings from RIS data cubes. Significant reduction in computation time, as well as a lack of tunable parameters makes this method desirable. • Spectral libraries have great use for visually comparing spectral features to help with pigment identification. Direct matching by algorithm is problematic because of normal variations in hue, scattering and the fact most paint layers are not optical thick enough in the near infrared. Knowledge of pigments and their distribution in paintings can inform conservators and scholars to better understand, and conserve these objects. In recent years, a variety of macro-scale non-contact imaging modalities have been applied to paintings to create classification/material maps. One modality, reflectance imaging spectroscopy (RIS) captures the diffuse reflectance spectral signatures of the artists' materials present in the objects which can be used for classification (by defining regions of the image cube that have the same spectral reflectance signature). In Cultural Heritage Science the ENVI spectral hourglass wizard (ENVI-SHW) has been used to create classification maps from RIS. This method is slow and requires an expert user input during the processing in order to find the reflectance spectra that define the spectral classes (spectral endmembers or exemplars). The pigments are then deduced from spectral features in endmember spectra and from other non-invasive analytical chemical analysis of sites defined by the RIS maps. This paper explores the potential of more automated algorithms, such as maximum distance (MaxD), that are based on the theory of convex hulls for doing these classifications as well as spectral matching algorithms to an optimal spectral library. While all of these algorithms were developed for Remote Sensing community, MaxD and the library matching examine all the spectra in the reflectance image cube, require little user input, and are reproducible and fast. The results from the analysis of reflectance image cubes (400 to 950 nm) of three paintings with MaxD and a spectral library matching algorithm are presented and compared to those obtained with the ENVI-SHW algorithm. The three paintings analyzed are the central panels from illuminated leafs, two from the Laudario of Sant'Agnese (c. 1340) painted by the Master of the Dominican Effigies, namely The Nativity with the Annunciation to the Shepherds and, Christ and the Virgin Enthroned with Forty Saints , and one cutting from a leaf from a Choir book, Saint Francis Receiving the Stigmata , painted by Cosmè Tura (c. 1470s). Specifically, the results include the spectral endmembers recovered from each method as well as the associated classification maps. The findings from this study show that the MaxD algorithm was able to find the majority of the reflectance spectral endmembers with little user input. [ABSTRACT FROM AUTHOR]
- Published
- 2020
- Full Text
- View/download PDF