This paper proposes a new framework for the eye centers localization by the joint use of encoding of normalized image projections and a Multi Layer Perceptron (MLP) classifier. The encoding is novel and it consists in identifying the zero-crossings and extracting the relevant parameters from the resulting modes. The compressed normalized projections produce feature descriptors that are inputs to a properly-trained MLP, for discriminating among various categories of image regions. The proposed framework forms a fast and reliable system for the eye centers localization, especially in the context of face expression analysis in unconstrained environments. We successfully test the proposed method on a wide variety of databases including BioID, Cohn-Kanade, Extended Yale B and Labelled Faces in the Wild (LFW) databases.