Ink Detection from Carbonized Herculaneum Papyri using Deep Learning

A Roman villa in the town of Herculaneum, next to Pompeii, housed a library with thousands of ancient Greek and Roman scrolls. The eruption of Mount Vesuvius buried this villa approximately 2000 years old ago which has been discovered. The scrolls can no longer be opened without breaking them because they were carbonized by the heat of the volcano. Since Herculaneum ink is carbon-based, it affords no X-ray contrast against the underlying carbon-based papyrus. Some detached papyrus fragments that can be read under infrared light can be used as ground truth data for a machine learning model that could identify otherwise undetectable ink from X-rays. The process of ink detection involves locating the inked portions of the papyrus using data from a 3D X-ray scan of the papyrus surface. The model has been trained by selecting a pixel from the binary label picture and sampling a subvolume from the surface volume at the same coordinates. In order to update the model weights, then backpropagate the known label data. Once trained, the model can be used to predict the appearance of a label picture from diverse input data. Three distinct models, namely U-Net, Attention U-Net, and a Pretrained Autoencoder with Attention U-Net, have been applied. Then, utilizing those three models, ensemble approaches have been employed (average, voting, and weighted average) to choose the optimum outcome. The best dice score of 0.766 has been achieved by the average ensembling technique. This project shows how machine learning can detect ink in 3D scans and decipher carbonized parchment contents.