Multi-light imaging for graphical heritage

Abstract

Historical artifacts in the category of graphical heritage are characterized by localized modifications of physical surfaces to convey textual or pictorial contents, such as manuscripts, drawings, engravings or reliefs. When affected by natural degradation or purposeful alterations, contents may be inaccessible with the naked eye or conventional digitization methods. Multi-light imaging involves capturing an object under various lighting conditions, encompassing multispectral imaging (variation in light spectra) and photometric stereo (variation in light directions). Multispectral imaging is used to record chemical surface variations resulting from inks and pigments, while photometric stereo captures variations in surface geometry and depth. By focusing on low-level operations such as image acquisition, calibration, shape reconstruction and visualization, the thesis caters to the needs of humanist scholars who require high-quality interpretable imagery for their investigations: A mobile image acquisition system is described, complemented by approaches for measuring and compensating longitudinal chromatic aberrations occurring in multispectral imaging, and heuristics for post-processing and calibration. In the field of photometric stereo, efficient light source configurations for mostly flat surfaces are evaluated and an analysis of the errors introduced by a simplified lighting model is conducted, which gives rise to a general error mitigation strategy. A case study in which poems are retrieved from indentations in paper demonstrates the practical applicability of these contributions. Another line of work is concerned with developing methods for evaluating the quality of graphical heritage imagery - in particular, with objectively assessing human text legibility in images. To this end, a novel dataset of manuscript images, rated for legibility by humanist scholars, is introduced. Being created with a novel study design, the expert ratings obtained are statistically validated. The dataset is then used to test potential quantitative estimators for text legibility in images.