Quantitative hyperspectral imaging (QHSI) is a non-destructive optical technique that provides accurate spectral measurements with high spatial resolution from virtually any surface. The high reproducibility of the calibrated measurements makes the QHSI technique also ideally suited for comparing different objects. By taking repeated measurements of the same object it can be used to detect changes, such as those caused by aging or conservation treatment.
The Nationaal Archief (National Archives of the Netherlands, The Hague) in cooperation with Art Innovation has been investigating the potential of QHSI for the analysis of archival documents. This poster discusses the workflows and processing techniques developed to enable an efficient analysis of the huge amount of data provided by hyperspectral measurements.
The SEPIA QHSI system in use at the Nationaal Archief measures 4 million calibrated spectral reflectance curves using 70 wavelength bands, in the 365-1100 nm range, covering a document area of 125 × 125 mm (5 × 5 in.). By covering this significant portion of the electromagnetic spectrum, from the near-ultraviolet to the near-infrared, it is possible to reveal features invisible to the naked eye, such as underdrawings, erased inscriptions, and media differences. The recorded reflectance curves are stored as pixel values in 70 grayscale images. Each image contains the reflectance of a document area at the corresponding wavelength band. This set of grayscale images creates what is known as the hyperspectral data cube of a QHSI measurement, which is over half a gigabyte of data.
The different data processing steps of the analysis workflows can make use of both the image and the spectral aspect of the hyperspectral data cube to extract and visualize the required information. For example, the readability of a palimpsest may be improved considerably by simply selecting a suitable near-UV or near-infrared spectral image. However, spectral data processing is typically required to achieve maximum readability or to map and quantify the amount of spectral changes such as discoloration of substrates and media caused by aging process.
For many applications it is important to distinguish among different classes of areas on the measured document (for example, to differentiate between media or to detect spectral variations caused by previous restoration). In order to generate a map from a hyperspectral data cube illustrating the locations of two different media, several analysis steps are required. During some analyses, it was found that applying a principal component analysis (PCA) to the spectral measurement data is a very effective first step in a workflow of vector calculus. The PCA method exploits statistical correlations within the hyperspectral data cube to condense all significant information into a much smaller set of images. Subsequent analysis steps, such as the comparison of the spectral features of different media and mapping of their distributions on the document, can then be carried out on this reduced image set. This speeds up the data processing considerably without sacrificing the accuracy or objectivity of the QHSI technique.