Compressed projections for localization of landmarks in the eye region

In this paper we describe a new system for localization of landmarks in eye region. Firstly, in the upper face area possible the eye regions are described by concatenations of integral and edge projections. The features dimension is reduced through Principal Component Analysis (PCA) technique, to accelerate the computation. The resulting features are fed into a properly-trained Multi-Layer Perceptron (MLP) that discriminates among various candidates. We test our approach on Cohn-Kanade database and we demonstrate that the system can achieve a considerable improvement in accuracy over state of the methods.