=============================================================================
	README file for the example files watch.xxx
=============================================================================


Description:	This network is an example of using neural networks
============	for tasks of image processing.

The network performs a 3x3 image convolution. The task was to detect
edges like the sobel operator does in conventional immage processing.


Pattern-Files:	watch.pat
==============	

The pattern file watch.pat contains one single pattern pair which
consists of two images: The input pattern represents an image of size
171x223 and the output pattern represents an image of size 169x221.
This is an example for a pattern file with two variable dimensions.
However, since there is only one pattern pair there is in fact no real
variety in size between different patterns. SNNS is able to cut small
parts out of variable sized patterns and to feed them into the network
during learning and test (see Hints:)

For easier understanding the two original images (input and output
image) are included in the SNNS distribution: watch_orig.pgm is a pgm
image file which shows a watch. watch_edge.pgm is the corresponding
edge image. You need a standard image viewer (like xv) to view these
images.


Network-Files:	watch.net
==============	

The network contains a trained network file with the following topology:
	9 input neurons (organized as 3x3 input mask)
	4 hidden neurons
	1 output neuron


Config-Files:	watch.cfg
=============

This network uses one 2D display in its standard configuration.


Hints:
======

To use this network and pattern file you need to define a sub pattern
scheme with the SNNS remote panel: You need to tell SNNS how to cut
sub patterns out of the image patterns. Open the sub pattern panel and
insert the following values:

		Input				Output
		Size	Step			Size	Step
dim 1:		3	5			1	5
dim 2:		3	5			1	5

You need to define a 3x3 input size since the network is also of these
dimensions. The step size of 5 is recommended to avoid long training
cycles during experimentation. In any case you should define equal
step sizes for the input and the output part of one dimension.
Otherwise the cut output sub pattern does not correspomd to the input
sub pattern and the network will not learn.

For more information about sub pattern handling please refer to the
user manual.


=============================================================================
	End of README file
=============================================================================