EDUCE: Imaging the Herculaneum Scrolls
Following the fragment study, Dr. Seales and his team traveled to the Institut de France in 2009 to perform the first-ever micro-CT scans of completely closed Herculaneum scrolls. These micro-CT images revealed much that was previously unknown about Herculaneum papyri and definitively proved it was possible to see, without physically opening the scrolls, every papyrus strand and fiber inside them, even the grains of sand trapped in tiny creases. This study provided the first ever volumetric images of a Herculaneum scroll, revealing the complexity of the internal structure and the utter scale of the task at hand. The scrolls were so tightly wrapped that in many places the scans showed little to no separation between layers, and the team learned much in their attempt to process the images.
First, naive application of existing segmentation techniques does not address the unique problems that Herculaneum scroll strata and that of other ancient manuscripts present. Depending on the type of material being scanned, scroll strata can appear fuzzy or almost indistinguishable due to time-induced distortion, disintegration, or other deformations of the writing medium. Some commonly used writing surfaces like papyrus can fray easily over time, while others like parchment made from animal hide often bubble or suffer holes. Such local defects in the surface make it difficult to visually follow a layer through the volume across the CT slices. Furthermore, undulations in and fusions of the scroll strata can cause the separations between layers to disappear and reappear at random, and tracking a single stratum through an entire scan becomes even more difficult as layers seem to merge together and then separate later on in the scan. The ink used to pen Herculaneum texts poses another challenge. Ancient writers used a mixture called “carbon black” which, when exposed to micro-CT scans, attenuates the x-rays the same way the carbonized papyrus, an carbon-based material on which it sits, does. Therefore, the ink of Herculaneum texts does not show up as readily in CT-scan data as that found in medieval texts, which usually contains metallic elements that are much denser than the writing surface and attenuate x-rays differently. In addition to these material-specific problems, micro-CT introduces its own anomalies and deformities into scan results, such as beam hardening, streaking, and ring artifacts, that must be compensated for in the final raw CT slices. These extra artifacts can make image processing much more difficult on scroll CT data and are heavily dependent upon the parameters selected at the time of the scan.
This confluence of challenges makes Herculaneum scrolls the most difficult digital restoration case encountered by Seales and his team. Following the 2009 micro-CT scans, the team spent the next few years prototyping a step-by-step computational approach for processing micro-CT data that would produce a digital representation of unseen text rich enough to enable textual scholarship, a process called “virtual unwrapping.” These advances address most, if not all, of the segmentation chjallenges presented by Herculaneum materials. Most recently, the team successfully developed and deployed a machine learning algorithm that is powerful enough to reveal the carbon ink that has proven so stubborn to researchers.
Innovative Machine Learning Tool Reveals Carbon Ink https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0215775)