NETWORK STRUCTURE
The project would involve use Kohonen's Self-Organizing Map (SOM) [1]. The network structure is as follows:-
Number of Inputs = 262144 ( 512 x 512 image)
Kohonen layer neurons = 5 (5 for gestures to be recognized)
The network Structure is shown below:-
This structure was developed for the optimized code.The original structure has another layer called Grossberg Layer ahead of Kohonen Layer, having structure:-
Inputs = 262144 ( 512 x 512 image)
Kohonen layer neurons = 5 (5 for gestures to be recognized)
Grossberg layer neurons = 4 (use binary code for counts, 0001 -1,0010-2,0011-3,0100-4,0101-5)
I had to change the structure so as to reduce calculation and get a more optimized algorithm for solving my problem.
The Kohonen Layer is based on the winner take all principal, which means that at a time only 1 neuron out of the 5 given fires(we assign output =1 for it). So for this problem the neuron which fires gives the count. Like if K2 fires then count =2, K3 fires then count =3 and so on.So the need of grossberg layer is eliminated.
The Kohonen Layer is based on the winner take all principal, which means that at a time only 1 neuron out of the 5 given fires(we assign output =1 for it). So for this problem the neuron which fires gives the count. Like if K2 fires then count =2, K3 fires then count =3 and so on.So the need of grossberg layer is eliminated.