Kornél Kovács1, András
Kocsor1, Csaba Szepesvári2
1Research
Group on Artificial Intelligence of the Hungarian Academy of Sciences
2Computer and Automation Research Institute of the Hungarian
Academy of Sciences
Face recognition is a highly non-trivial classification problem since
the input is high-dimensional and there are many classes with just
a few examples per class. In this paper we propose using a recent
algorithm - Maximum Margin Discriminant Analysis (MMDA) - to solve
face recognition problems. MMDA is a feature extraction method that
is derived from a set of sound principles: (i) each feature should
maximize information transmission about the classification labels,
(ii) only the decision boundary should determine the features and
(iii) features should reveal independent information about the class
labels. Previously, MMDA was shown to yield good performance scores
on a number of standard benchmark problems. Here we show that MMDA
is capable of finding good features in face recognition and performs
very well provided it is preceded by an appropriate preprocessing
phase.
