🚀 go-pugleaf

This is a summary of the replies I recieved to my question. Sorry for it being 
late. I had been away for a while. My original question is shown below.

>Has anyone used or come acros papers on neural nets for image  processing. 
>This is not the same as using neural nets for pattern recognition. I was
>wondering if it is possible  to use neural  nets for image enhancement, 
>edge detection and segmentation of objects in an image.

From: sjreeves@eng.auburn.edu (Stan Reeves)


For image restoration, see:

@Article{paik92,
  author =  "Joon K. Paik and Aggelos K. Katsaggelos",
  title =   "Image Restoration Using a Modified Hopfield Network",
  journal = SP,
  year =    "1992",
  volume =  "40",
  number =  "2",
  pages =   "1-41",
  month =   "February",
  OPTnote =       ""
}


From: sjreeves@eng.auburn.edu (Stan Reeves)
I inadvertently posted an incorrect reference on image restoration using
neural networks.  My info was based on a preprint.  It turns out that the
paper was moved from IEEE Transactions on Signal Processing to IEEE 
Transactions on Image Processing when the new journal got started.  I 
apologize for the error.  Here's the correct reference:

@string{IP = "IEEE Transactions on Image Processing"}

@Article{paik92,
  author =  "Joon K. Paik and Aggelos K. Katsaggelos",
  title =   "Image Restoration Using a Modified Hopfield Network",
  journal = IP,
  year =    "1992",
  volume =  "1",
  number =  "1",
  pages =   "49-63",
  month =   "January",
  OPTnote =       ""
}

Sorry for the inconvenience.


From: SCHOLAR@Meena.CC.URegina.CA
Sorry I can't remember the name of the book, but I saw one in the massive
computer section of Codys Books in Berkeley that had an example of the
use of a neural net to make line drawings of pictures.  The network was 
based on some theory of about human perception and the resulting drawing
looked much more like an artist's rendition than like something computer
generated.  The book must have been on digital image processing, because
there was only a small blurb on this technique and a few pictures, comparing
this technique to more standard ones.
			Joshua Scholar (scholar@meena.cc.uregina.ca)

From rak@crosfield.co.uk Mon Aug 24 18:29:29 1992
No pointers, I'm afraid, but we were thinking of trying the same thing.
We need a smarter sharpening algorithm for a scanner that can be taught
the difference between film grain and real features to reject one and
find the other, and things like that.  We have found no such references but
we haven't looked very hard either.  Our trouble is we are only just 
starting to get into neural nets.

Crosfield makes scanners for the printing industry.  If we can help, 
please let us know.


From esmythe@fergie.atl.ge.com Mon Aug 24 20:09:01 1992
Hecht-Neilson was given a tidy sum of money from SDIO (Strategic
Defense Initiative Office, the people who fund "star wars") to attack
the frame-to-frame correlation problem.  This was used for enhancement
using _multiple_ images, but may not be the thing you want.

I haven't seen anything of it, though, so it: a) was so good
that the government classified it, b) failed and nobody is talking
about it, or c) I missed it.

This probably doesn't help much.

cheers,

From fred@dip1.ee.uct.ac.za Tue Aug 25 15:12:46 1992
Last year I did some research into using backpropagation networks for
edge detection. I have not had a chance to do much followup work but 
the preliminary results that I achieved looked quite promising.

One of the things I did was to train a network on *ideal* edges. These
were just three by three kernels and the output indicated whether the
center pixel in the kernel was an edge or not.

             0 0 0              0 X X
             X X X              0 0 X
             X X X              0 0 0

             Edge               Not an edge

I trained the network on about thirty ideal cases like that and then
tested it on some real data. For a network trained with data like this
I was surprised at the results - very promising. I have been meaning
to go back and try this with a larger kernel size and more realistic
training data but I haven't had the chance as I am now investigating
neural networks for pattern recognition in images.

Hope this helps you.

Fred


From @shannon.ee.wits.ac.za:ian@concave.cs.wits.ac.za Tue Aug 25 17:38:25 1992
Our deptmartment has produced a couple of technical reports on Neural Nets and
edge detection which were also presented at the SA Symposium on Pattern 
Recognition in Dec last year. The work is on detection and not on recognition
I can send you copies if you like.

From @durham.ac.uk:des3mjj@capella Tue Aug 25 18:25:52 1992
Yes, I went to the IEE conference on Image Processing and its Applications
in Maastricht, Netherlands, at Easter. There was a paper presented which
might interest you, given by a guy from King's College, London, UK. 

He had used a neural network to do edge detection on an image of light and
dark circles, basically by telling it the right answer and training it a lot. 
There was some success - he didn't claim that his technique was better than
conventional methods, he did say that it was sometimes faster to apply 
his technique to new situations because little human intervention is needed
in the training process, cf. Gabor filters etc. which might need to be
manually fine-tuned for the situation.

It is on pp226-229 of the IEE proceedings for this conference. I have a copy
here, if you send me your snail-mail address I can send you a photocopy if you
like. In return, can you keep me up to date with any successes you might have
using this kind of idea? I'm interested in it too, although my work at the
moment concerns Fourier Transforms and things.