Fw: [ai-neuralnet-backprop] respected sir

Greetings All,

Recently this email, (Below) came to a list I host. I really don't know much
about fourier descriptors, and I thought that possibly someone in Perl-AI
might know more about this gentlemans subject than I. If anyone can help
with the below message, I'm sure it would be greatly appreciated.

~ Josiah Bryan


P.S. If anyone replies, could you CC it to Perl-AI and/or
[email protected]?
Thankyou!


----- Original Message -----
From: vikram ramaswamy <[email protected]>
To: <[email protected]>
Sent: Saturday, January 20, 2001 8:02 AM
Subject: [ai-neuralnet-backprop] respected sir



   Respected Sir,

   I am vikram an undergraduate student from india. I am currently working
on a
   project on shape classification using a feed forward neural network
trained by
   standard backprop algorithm.

   I initially detect the edges of the shape to be classified and find the
   fourier descriptors(fds) of the edge image. I use these fds as the input
to
   the neural network.

   This approach has been followed before and a paper in this topic has been
   published. We are infact trying to implement what the below listed
authors
   have done before (only that the objects to be classified are different):

   Hongbong Kim, Kwanhee Nam, 'Object Recognition of one DOF by
back-propagation
   neural net', IEEE transactions on Neural Networks Vol.6 No.2, March 1995.



   I have a few doubts about the  fourier descriptors. Can you please kindly
   excuse the trouble and answer my questions?

   In my study , I have used only 16 fds as mentioned in the above paper.

   1) I take an edge image of an (eg. circular) object and find its fourier
   descriptors. I now rotate the original image , detect the edges and get
the
   fourier descriptors(fd) of this image. It is mentioned that fd's are
   insensitive to rotation. Does this mean the 2 sets of fd's mentioned
above
   must be identical? If not, what is the way in which the fd's of the
rotated
   images related?

   2)Also, when we use an object of a different shape eg. rectangle, and get
the
   fd's , I would expect that fd values be drastically different from the
values
   obtained for the circular object. Is this assumption justified? Moreover,
I
   took a circular object got its fd's; rotated the object obtained another
set
   of fd's. I computed the difference between the two sets of 16 fd . I then
   found the difference between the fd of a circular object and one for a
   rectangular object. I am troubled by the fact that the 2 differences are
   comparable. I thought (fd for circle) - (fd for rotated img. of circle) <
(fd
   for circle) - (fd for rectangle).

   By difference I mean foll.:  we calculate 16 fd. take (1st fd value for
circle
   - 1st fd value for rotated image of circle). This is done for all the 16
   values.

   Max difference between circles must (i guess)be < Max difference between
   circle and rectangle.

   Since this is not so, can you kindly clarify this situation?


   IF ANY OTHER APPROACH IS POSSIBLE FOR SHAPE CLASSIFICATION USING FEED -
FWD
   NEURAL NETS KINDLY INFORM ME OF THE SAME.

   Am anxiously awaiting your reply,

   Yours respectfully,

   vikram


   To unsubscribe from this group, send an email to:
   [email protected]