| |
 |
|
| Computers Forum Index » Computer Artificial Intelligence - Neural Nets » Convolutional network: How to find relevant input... |
|
Page 1 of 1 |
|
| Author |
Message |
| ChristophW... |
 Posted: Sat Sep 05, 2009 11:38 am |
|
|
|
Guest
|
Hello all,
recently I started experimenting with neural networks, specifically
convolutional networks for character recognition. The images contain
lots noise from poor scanning. Sometimes the NN would not deliver a
definite answer but multiple candidates so I am thinking about using
other verification methods afterwards.
In cases when multiple outputs are 'high', is there a method to go
backwards through the activated NN for each character to find out,
which parts of the image the NN believes to be the respective
character? If this explanation is confusing, picture that for each
candidate I would like to highlight the part of the input image that
represents the character.
I have read that by using a special training method for convolutional
neural networks the topmost hidden layer will show a stylized version
of the character. I have not tried this yet but I believe that this
will not help me anyway because I need to find the outline of the
character in the input, which could be slightly translated or
distorted in addition to the noise.
Thanks |
|
|
| Back to top |
|
|
|
|
|
All times are GMT
The time now is Mon Nov 23, 2009 1:34 pm
|
|