By modelling the noise in training sets and in the image, and by removing outliers in the data, we have made improvements to subspace based face recognition systems. See <a href="http://users.monash.edu.au/~tchin/">TJ's Homepage</a> <p>