papers available on high performance OCR

[Lucas S M]

  Summary:

 The following two papers discuss recent work on applying
 scanning n-tuple classifiers to handwritten OCR.  The 
 first is a journal paper which gives some background
 and all the technical details.  The second is a paper for
 a forthcoming conference which includes more up-to-date
 results and more detailed timing analysis.

 The main feature of the method is the incredible speed.
 If we ignore the pre-processing time, we can train the
 system at a rate of over 20,000 character images per second,
 and recognise about 1,200 characters per second, on 
 a humble 66mhz Pentium PC.  If we include pre-processing
 time, then we can still train (recognise) 500 (200) chars per second.
 The fast training and recognition speeds allow the system
 parameters to be optimised very quickly.

 Best accuracy reported is 98.3% on the CEDAR hand-written digit
 test set.  This is not quite at good as the best reported in
 the literature for this data (98.9%, to the best of our knowledge),
 but offers a significant speed advantage.

 The following two papers discuss recent work on a simple unified
 approach to evolving ALL aspects of a neural network, including
 its learning algorithm (if any).  The first uses a grammar based
 chromosome, the second uses a set-based chromosome.  The latter
 approach appears particularly promising as a method of part designing/
 part evolving neural networks.

-----------------------------------------------------------------------