Main  ◊  Intro  ◊  Active Contours  ◊  Locating Eyes  ◊  Hough Transform  ◊  Neural Networks  ◊  Network Training  ◊  Applet Intro  ◊  The Applet  ◊  Links  ◊  Home Page

Training for Face Recognition

Training a neural network via back-propagation using an image has been done before.   The actual pixels can be fed into the network as the inputs.   This approach works great when trying to recognize textures or objects with fixed orientation and scale.

This method is not so practical for face recognition since the face is never a constant and there are other variables such as orientation, lighting, expression, distance to camera, etc.  Also for images with large numbers of pixels the computation can become very complex.  For example a 200x200 image would require 40000 inputs!

I choose instead to parameterize the face using information from the contour and eyes found earlier as well as some picture information.

I broke the face into 8 parameters

1. Curvature of the face
2. Gradient energy of the face
3. Perimeter of the face / Area of the face
4. Width to height ratio
5. Distance between the eyes relative to the width of the face
6. Position of the eyes relative to the height of the face
7. Average gradient of the face
8. Relative size of the eyes

I then used these parameters to train the neural network for a set of different faces.