Spectral unmixing as a preprocessing step for SVM-based material identification in historical manuscripts

Abstract

When performing material identification from hyperspectral images, a common challenge is the mixing of spectral signals at boundaries between materials. This study investigates spectral unmixing as a preprocessing step to improve machine learning-based classification of inks and writing supports in documents. Hyperspectral data of mock-ups and historical samples were acquired in the VNIR and SWIR ranges, including metallo-gallate, carbon-containing, and non-carbon-containing inks (sepia or mixtures with iron gall) applied to paper and parchment. A subtractive mixing model with automatic endmember extraction was used to generate presence maps and exclude pixels below a concentration threshold. Three support vector machine classifiers were trained using (1) unprocessed reflectance spectra, (2) reconstructed spectra from unmixing, and (3) pure unmixed spectra. Reconstructed spectra provided the best overall performance and classification maps, while unmixed spectra outperformed in ink identification, particularly bleed-through detection. Unmixing also revealed areas of lower classification confidence, offering potential for broader hyperspectral applications.