On Wed 14-Jun-2017 at 15:07:46 +0200, "Arthur B." <[email protected]> 
wrote: 
> I also would like to try the "Artificial Neural Network Classifier" but i'm a 
> little overwhelmed by the quantity of settings to be adjusted, and have 
> difficulties in taking out the head of documentation (which is a little 
> tricky without any artificial neural network knowledge). I would especially 
> like
> to know quickly if this classification can suit our objectives, and default 
> settings aren't enough to run correctly the processing. Can you help me with 
> the esential elements to be parametrized to have basic but correct results 
> (depending on the image, dataset or expected results i
> guess..) ?
>

Hi,

For most parameters, the default values are OK. Since OTB uses OpenCV
for the ANN implementation, you can get more information here:

http://docs.opencv.org/2.4/modules/ml/doc/neural_networks.html

The only parameter you need to really play with at the beginning is the
network architechture (how many neurons per hidden layer). This is done
using:

-classifier.ann.sizes        <string list>    Number of neurons in each 
intermediate layer  (mandatory)

The input layer will have a number of neurons equal to the number of
input bands. You don't need to give this value on the sizes list. The
ouput layer has always as many neurons as the number of classes of your
problem. This is automatically found by the application. So for a
network working on 4 band images and 6 classes, if you choose 2 hidden
layers with 5 neurons per layer, the network architecture is represented
by:

4 5 5 6

but you only have to give the hidden layers to the application:

-classifier.ann.sizes 5 5

In the GUI application or in Monteverdi, you click twice on "+" and set
5 and 5.

I would suggest starting with one single hidden layer and as many
neurons as classes and proceed from there by adding neurons or layers.

However, bear in mind that for typical image classification problems,
this kind of ANN (multi-layer perceptrons) will give you worse results
than Random Forest or SVM.

Best wishes,

Jordi



> Thanks a lot !
>
> Arthur
>
> -- 

-- 
-- 
Check the OTB FAQ at
http://www.orfeo-toolbox.org/FAQ.html

You received this message because you are subscribed to the Google
Groups "otb-users" group.
To post to this group, send email to [email protected]
To unsubscribe from this group, send email to
[email protected]
For more options, visit this group at
http://groups.google.com/group/otb-users?hl=en
--- 
You received this message because you are subscribed to the Google Groups 
"otb-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
For more options, visit https://groups.google.com/d/optout.

Reply via email to