From the Archimedes' Palimpsest to the Vercelli Book: Dual Correlation Pattern Recognition and Probabilistic Network Approaches to Paleography in Damaged Manuscripts

The human capacity to visually discern and detect information is remarkably powerful. This ability forms the basis for the field of paleography, specifically in terms of text identification. However, while traditional paleographical methods yield impressive results for data characterization, a mathematical approach to identification is desirable as a means of additional verification and as a basis for comparison between varying proposals of text identities, especially in the case of damaged manuscripts in which text is wholly or partially illegible. Executed judiciously, mathematical approaches can aid the transcription efforts for manuscripts and encourage discussion of text identity with appeal to statistical distribution, based on both v!isual and contextual information. !

Derek Walvoord initially employed such an approach for the Archimedes’ Palimpsest in 2004.1 He developed a correlation recognition program and probabilistic network system for data analysis that utilized both visual and contextual information in order to create character identity distributions for the data of interest. I am currently working to adapt and improve his approach in order to analyze image data for the Vercelli Book. The oldest existing record of Anglo-Saxon poetry and homilies, this 10th-century manuscript was imaged in March 2013 by a team from the Lazarus Project at the Archivio Capitolare in Vercelli. Led by Dr. Gregory Heyworth, the team was composed of Dr. Roger Easton, Kenneth Boydston and Dr. Keith Knox, accompanied by several students from the University of Mississippi, including myself. Utilizing reflective and transmissive LED arrays in 12 wavebands ranging over the near-ultraviolet, visible, and near- infrared regions, we imaged the manuscript with the aim of penetrating the chemical reagent a!pplied in the 19th century that obscures sections of the text. !

In order to determine the identities of characters and character fragments which image rendering do not conclusively reveal, I am designing a neural network system which will make use of correlation filtering to classify the data, similar to Walvoord’s system; however, by utilizing subsampling of the initial and subsequent output layers, I will be able to meaningfully analyze fine details of the data in a way that Walvoord’s system did not allow. Moreover, my updated system will be able to weight the importance of particular features in determining class membership via a learning architecture, such that each training image improves the classifier function. This method should overcome problems traditionally associated with correlation filtering methods such as blur, lighting differences, and lack of contrast because of its subsampling and weighting mechanisms. Depending upon the results of this neural network approach of classification for the Vercelli Book data, I will evaluate whether employing a probabilistic network system similar to that used for the Archimedes’ Palimpsest will significantly i!mprove overall data classification and proceed with research accordingly.!

This system would serve to bridge the gap between our knowledge of the Vercelli Book’s contents and the information available to us, providing a probabilistically accurate transcription of data associated with a piece of the humanities corpus that has been obscured from view for over a century. This invaluable data will contribute place and context to our current conversations considering Anglo Saxon literature, language, and culture, as well as manuscripts in general. Moreover, this system will ultimately serve to empower paleographers and computer scientists alike by exploring the relationship between the visual acuity and classification knowledge associated with traditional transcription efforts and neural network-based approaches for mathematical ones.

References
Walvoord, Derek J. (2008). “Advanced Correlation-Based Character Recognition Applied to ! the Archimedes Palimpsest.” PhD diss., Rochester Institute of Technology.

Sermanet, Pierre et al. (2012) “Convolutional Neural Networks Applied to House Number Digit Classification.” Paper presented at the 21st International Conference on Pattern Recognition, Tsukuba, Japan, November 11-15.