Accurate Eye Localization under Large Illumination and Expression Variations with Enhanced Pictorial Model

As the first step in a face normalization procedure, accurate eye localization technique has the fundamental importance for the performance of face recognition systems. One of the most classical methods to address this is the pictorial model where the appearance model and shape constraints are optimized together. However, under extreme illumination changes and large expression variations, the simple Gaussian appearance model and the localization-based shape constraints used in the pictorial model are not capable to handle the complex appearance and structural changes appeared in the given face image. In this paper, we enhanced the pictorial model by combining the strength of illumination preprocessing, robust image descriptors, probabilistic SVM and an improved structural model which are invariant to scale, rotation and other transforms. Experimental results on CAS-PEAL dataset demonstrated that the proposed model can accurately localize eyes in spite of large illumination and expression variations in face images.