UNIVERSITY OF HERTFORDSHIRE
            COMPUTER SCIENCE RESEARCH COLLOQUIUM

                            presents

          "High Accuracy Recognition of Connected Script
               Languages in Old Printed Books"

                     Professor William Clocksin
                    (School of Computer Science,
                   University of Hertfordshire, UK)

                    8 December 2010  (Wednesday)
                       Lecture Theatre A161

                  Hatfield, College Lane Campus
                            1 -2  pm


                  Everyone is Welcome to Attend

                  Refreshments will be available

Abstract:

  I will describe Quroyo, a high performance system implemented in
Java for the recognition of Syriac, Arabic and Mandaic script. The script
sources are scans of old printed books that were typeset by hand
between the 17th and early 20th centuries. These sources are plentiful -- 
some 20,000 of these books are in the British Library alone -- and 
scholars would like to have machine-readable transcriptions to aid 
research.  Unlike previous optical character recognition systems, Quroyo 
is not based on a trainable classifier. No statistical language model is 
used, nor is a training set required. Instead, exemplars model the 
graphic appearance of each character in the alphabet. A contour matcher 
based on dynamic space warping finds the candidate matches of examplar 
characters to deformed characters within each connected group of pixels. 
In connected script languages, the baseline extender provides important 
but unreliable context information.  Without requiring any segmentation 
information, a knapsack decoder (similar to a Viterbi decoder) then 
finds an optimal sequence of characters within each connected group, and 
then words are detected by maximisation of spatial expectation. The 
numerous diacritic marks used in the target languages are decoded by an 
efficient hypergraph matching algorithm.  Except for a few cases 
(approximately 1 in every 2000 characters) where characters are 
seriously broken by typesetting faults, the character recognition 
rate is 100% at a speed of about 4 characters per second. The system 
has been tested on hundreds of pages of Syriac in all three of its 
ancient script forms, and on scores of pages of Modern Standard Arabic.

---------------------------------------------------
Hertfordshire Computer Science Research Colloquium
http://homepages.stca.herts.ac.uk/~nehaniv/colloq