Volumetric Texture Analysis for Unsupervised Ink Detection in Herculaneum Papyri

The Herculaneum papyri, carbonised by the 79 AD eruption of Mount Vesuvius, can be imaged by X-ray micro-CT, yet their carbon-based ink is nearly iso-attenuating with the carbonised substrate and usually leaves no hight contrast in the reconstructed volume. State-of-the-art ink readers are supervised neural networks trained against labour-intensive labels and tuned on cropped regions of each segment; their letter-by-letter legibility is uneven and they offer little explainability. Motivated by the community’s call for a clearer understanding of the patterns underlying ink detection, we present a deterministic, training-free alternative that treats the unwrapped depth slab as a genuine 3-D signal. Three established texture descriptors — first-order statistics, Laws’ texture energy and block spatial autocorrelation — are lifted to volumetric form and evaluated both individually as stand-alone ink responses and jointly as a per-voxel mean fusion; the resulting volumetric response is projected to 2-D under a per-specimen depth prior, smoothed by a soft Markov random field and binarised at an adaptive per-specimen threshold. Across 25 manually annotated Greek-letter specimens of Scroll 5 (PHerc. 172), under a δ-pixel exclusion band around the ink-mask contour, the best single descriptor reaches a mean Matthews correlation coefficient (MCC) of 0.374, strictly positive on every specimen. Run on the full segment, the sameoperator also surfaces partially recognisable text whose letters match the community ink-prediction of a neighbouring segment that overlaps with the analysed segment in volume. The pipeline is thus offered as a deterministic, explainable complementary signal that, in favourable cases, supervised neural detectors may consume as an additional input or weak label.