I have to write a neural net for my uni course to recognise peoples faces. The pictures may be at different orientations to those given as training data.
Does anyone any ideas on what network would be best (I was thinking back propogation) and algorithms to preprocess the image to orientate it.
I inted to scale the pricture so that the actual face is always the same percentage of the whole image before feeding the pixel values into the net.
The only way I can think of to find the orientation is to rotate the input points one unit at a time, monitoring the output to find the most likely image. This would seem to have a very large processing overhead though. Any thoughts?