||Sala conferenze, al Disi, num 322 - 3° piano
||Robust and precise eye localization for face registration and recognition
||Dott. Raffaella Lanzarotti
||Universita' degli Studi di Milano
||The current challenge for Face Recognition (FR) systems is to find
solutions adequate to real-world applications, that is completely
automatic and robust in unconstraint contexts.
In this framework, one of the critical unsolved problem regards
the precise and automatic localization of the facial features
(eyes, nose and mouth): most of the FR methods (e.g. PCA, ICA,
LDA, LFA, LNFM, LS-ICA, Kernel-PCA, Kernel-ICA) require either to
warp the input images, or at least to normalize them. This makes
the facial feature localization particularly crucial, above all
considering that even small localization errors make the
recognition process fail.
In this seminar, we present our eye localization method which
consists of two steps: at first a face detector localizes roughly
the position and extension of the faces represented in the image;
then the eye detector processes the found sub-images to localize
precisely the eyes. Here we focus on the second step that is built
by modelling the eye at different level of details by means of the
Haar wavelet transform. Two Support Vector Machines are trained by
examples having as input a subset of the Haar wavelet coefficients
selected to optimize a trade off between accuracy and efficiency.
The first classifier aims to roughly localize the eye within the
face, the second aims to refine the localization, determining a
precise position of the eye center.
We conclude the seminar reporting the performance of our algorithm
on standard databases (XM2VTS, BANCA, FRGC, FERET, and BioID) and
comparing it with the best algorithms presented in the literature.