garydean@images.cs.und.ac.za (04/26/91)
I'm currently studying for my Computer Science Honours and would like to use neural nets to solve the problem of digit recognition. I plan to have a grid as input and would like to train the network using patterns of the digits 0-9. After training, I would like the network to distinguish a digit entered albeit in a slightly distorted form. I have been reading the volumes available from the PDP research group. I have tentatively decided to use back propogation but would like any form of comment or references to help me. Thanx in antcipation. Gary Nicholson. Reply to : garydean@images.CS.UND.AC.ZA
arms@cs.UAlberta.CA (Bill Armstrong) (05/02/91)
garydean@images.cs.und.ac.za writes: >I'm currently studying for my Computer Science Honours and would like to use >neural nets to solve the problem of digit recognition. ... >I have been reading the volumes available from the PDP research group. I have >tentatively decided to use back propogation but would like any form of comment >or references to help me. ... >Gary Nicholson. I have used adaptive logic networks for OCR. They were tested on the Highleyman data from the US Post Office, which had handwritten numerals 0 - 9, as you intend to use. The logic networks proved to be quite immune to salt-and-pepper noise and rotation of synthesized characters, so I'm sure you would have no problems in making an OCR system with them. I suspect the system would be faster than a backpropagation network both for learning and execution. The code is available by ftp from menaik.cs.ualberta.ca [129.128.4.241] in pub/atree.tar.Z. Here is a reference with some experiments on noise immunity and rotation, done with a less powerful early adaptive algorithm. W. Armstrong and J. Gecsei, "Adaptation Algorithms for Binary Tree Networks", IEEE Trans. on Systems, Man and Cybernetics, 9, 1979, pp. 276-285. -- *************************************************** Prof. William W. Armstrong, Computing Science Dept. University of Alberta; Edmonton, Alberta, Canada T6G 2H1 arms@cs.ualberta.ca Tel(403)492 2374 FAX 492 1071
jondarr@macadam.mqcs.mq.oz.au (jondarr c g h 2 gibb) (05/09/91)
This is sort of for Gary, but if he doesn't get it, maybe someone else can
help.
I can't work out the address of images.cs.und.ac.za - it's that simple.
Here I am in the land of oz, an honours student doing benchmarking for the
common neural network simulation algorithms on the standard problem set
( which I'm expanding as I go ), and I wondered if there was anyone else
out there doing this sort of thing, etc, etc.
I have some interest in digit or character recognition, and intend to include
these in my benchmarking. I have collected oodles of references published
on net and referred to from there on in. I have toyed with algorithms of
standards ranging from basic implementations of Lippmann's interpretations
and simplifications through to 'real' implemented programs.
If anyone has any useful references on benchmarking ( aside from the standard
Fahlman, Veitch, Ackley, etc ) which actually do fair comparisons, and do not
just promote their own pet algorithms, would they be kind enough to post or
e-mail where such items can be found.
I'm probably not the only one with this sort of interest.
Thanx, etc,
jondarr
[ jondarr@macadam.mqcs.mq.oz.au - 137.111.160.57 ]
P.S. anyone else out there in the land of oz ?