Durham Priory Library OCR Project

Thanks to a recent collaboration with the Durham Cathedral's Priory Library, Rescribe has completed a project creating OCR models with an output for four medieval manuscripts of over 90% accuracy rate. The manuscripts used were:

• MS A.II.4: Bible of William of St Calais
• MS B.II.16: Augustine, In Johannis Evangelium
• MS B.II.18: Augustine, Sermones de verbis Domini et Apostoli
• MS B.III.1: Origen, Homilies on Old Testament

We plan to release the models we created, as well as the supporting software and data, very soon, on LatinOCR.org.

The Priory Library supplied Rescribe with high-resolution images of the chosen manuscripts, and Rescribe used the images to train OCR models and machine-readable output for each manuscript.

The project had several objectives at its core: firstly, we wanted to resort to using exclusively open source tools to train an OCR model for the manuscripts at hand; secondly, we were working on as time-efficient a model-building process as possible, finding the minimum amount of ground truth necessary for an effective model; and figuring out a process to expand abbreviations and glyphs in medieval manuscripts as accurately as possible.

The program of choice for the project was OCRopus, an open source program with extreme adaptability for a variety of purposes. The single steps – binarization, segmentation, training, recognition – are all divided up and can be altered individually. Since we used different open source image preprocessing programs – ScanTailor and Gimp – for our pre-processing purposes, we left out OCRopus’ binarization step, for example. Since the original OCRopus did not contain all the Unicode characters used for medieval manuscripts, we adjusted both training and recognition script to include the same. The complete character set used for the occasion was the following:

[ ~ ! & ' ( ) * , . 0 1 2 3 4 5 6 7 8 9 : ; < = > ? A B C D E F G H I J K L M N O P Q R S T U V W X Y Z [ a b c d e f g h i j k l m n o p q r s t u v w x y z { } § « ¶ · » Ã Æ Õ ß à â ã æ è ê ì î ò ô õ ÷ ù û Ā ā đ Ē ę ē Ĩ ĩ Ī ī Ō ō Œ œ Ũ ũ Ū ū ǣ ǽ Ẽ ẽ ‚ † ‡ ‹ › ※ ⸗ ꝑ ꝓ ꝗ ꝙ ꝛ ꝝ ꝫ ꝯ ꝶ ] ꝗͣ r̄ t̄ pͣ tͣ qͥ qͦ gͦ qͣ mͦ uͦ Vͥ pͥ nͥ gͥ n̄ p̄ mͥ qͣ p̄ c̄ tͥ m̄ q̄ x̄ s̄ n̄ ł ƀ ľ ę ħ ƀ ľ ł ħ ƀ

The initial output of the OCR recognition process preserved these characters. We further processed the documents in two steps: by first expanding “non-standard” formations; i.e. xpm, ihs, eccla, gra; in a second step, the "standard-expansions" where added. This means that ambiguous abbreviations like ā could be expanded to either "am" as a standard abbreviation, or "an" or something entirely different, context-related, as a non-standard abbreviation.

For the three manuscripts displayed, following models were trained:

• MS A.II.4: 4 pages (304 lines)
• MS B.II.16: 4 pages (139 lines)
• MS B.II.18: 2 pages (157 lines) + MS B.III.1: 2 pages (150 lines) = 4 pages (307 lines)

By combining the ground truth for MS B.II.18 and MS B.III.1, better results with fewer pages of ground truth could be achieved; the combination of MS B.II.18 and MS B.III.1 with MS A.II.4, however, did not yield a functional model for MS A.II.4, so an individual model was built for that purpose.

The results ranged within a character error rate of ca. 0-10%, meaning, the resulting pages had a word accuracy of roughly 90% and more.

OCR Samples

Durham Cathedral Library MS. B.II.16 - Augustine, In Johannis Evangelium, page f.4r

OCR Results

OCR Results (expanded abbreviations)

Durham Cathedral Library MS. B.II.18 - Augustine, Sermons on the Gospels, page f.5v

OCR Results

OCR Results (expanded abbreviations)

Durham Cathedral Library MS. B.III.1 - Origen, Old Testament homilies, page f.3v

OCR Results

OCR Results (expanded abbreviations)

Durham Cathedral Library MS. A.II.4 - Bible of William of St Calais, page f.14r

OCR Results

OCR Results (expanded abbreviations)