Training a neural network for face recognization

 

I'll explain here how I created and trained the neural network that is given in the demo with my pattern matching neural network editor.

The first step is to create a new pattern matching neural network (File->new)

On this dialog we can define the input window size (in pixel) of the neural network and the input mask. pixels in black on the mask won't be given in the input vector of the neural network. Here it is usefull to ignore the background behind the face.

As We had now difined the input of the neural network we can create the neural network by cliking on "New" in the neural network panel :

I do not know the optimal structure of neural network to use so this is just an example (but a working example :-)).

As the network is now created we can start to perform learning. To begin learning, I've cut about 15 face image of 25x25 pixels to load them as matching examples and about 30 not face image to load them as not matching examples.
Then start learning process :

The neural network learned the samples I had given quite quickly. The I tried to use it to find face on various images comming from internet and I added matching and not matching images as the network was doing mistakes.

At the end the neural network given in the demo has been trained on about 100 matching samples and 400 not matching images.

I'hope my little you understand a bit better how does it works, for any question mail me.