Analysis of Multispectral Imaging of Illuminated Medieval Manuscripts

Abstract

Over the past few years Multispectral imaging and Hyperspectral imaging have gained popularity in various fields including remote sensing [1], forensic analysis [2], and arts and painting analysis [3]. In this research, we are focused on multispectral images of illuminated manuscripts. These are composed of natural mineral and plant pigments applied to a substrate made from animal skin, using various natural binders, as well as gold leaf. We start by analysing modern ground-truth data, where modern versions of the medieval pigments have been applied to paper. This allows us to explore if the variations seen can be explained using simple linear models. The research then advanced to performing basic analysis of identifying the pigment used in the multispectral images of medieval illuminated manuscripts provided by The John Rylands Library. We applied the commonly used algorithms Spectral Angle Mapper (SAM) and Spectral Correlation Mapper (SCM) (see Chapter 6 for reference), and the applicability of the linear model has been directly verified using fortuitous areas of accidental pigment overlap in the illustrations of the illuminated manuscripts. We then applied distance-based algorithms Euclidean and Mahalanobis distance (refer to Chapter 7), and then doing further analysis by using Principal Component Analysis (PCA) (refer to Chapter 8). Lastly Kullback-Leibler divergence (KL-divergence) (refer to Chapter 9 and 10) on these manuscript images has been performed. The results obtained by applying SAM and SCM shows that SCM outperforms SAM in almost all the documents in the illuminated manuscript database at hand, whereas satisfactory results have been obtained by Euclidean and PCA. However, Mahalanobis distance does not work well with our illuminated manuscript images. KL-divergence out-performs all of the applied algorithms, and various 3D structures can be seen in these results obtained including visibility of pores, overleaf impressions, 3D structure in the text as well as cracks and highlights in the gold region, which is not properly seen in the results obtained by applying other algorithms. Overall, we show that KL divergence allows the visualisation of effects not seen with any other measure, and hence a deeper and more specific understanding of the chemical and physical details of the manuscript. Many of the observations we were able to make were, as far as we know, unique. In future work, ground-truth data such as x-ray studies or 3D laser scans would be needed to fully verify our results. Apart from these experiments additional experiments have been performed on the “Latin MS 30 Nicholas of Lyra Commentary on the Bible 123v” when the image was captured using updated lighting panels in Chapter 11. These results have been quantified using Jaccard similarity index with a given ground truth image to show which algorithm performs better when the results are compared to the ground truth. Even though SAM and SCM perform better according to the similarity index, various 3D structures (as explained in chapter 10) can be seen while applying KL-divergence on the manuscript image. The colour images of the manuscripts in the multispectral imaging dataset as well as the various digital collections, which are part of the Manchester Digital Collections can be viewed online and downloaded at https://www.digitalcollections.manchester.ac.uk/

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Timothy Cootes (Supervisor) & Bijan Parsia (Supervisor)