Robust face authentication can be achieved by fusing the output of three independent experts, which are implemented by means of SVM classifiers. Each of the experts is trained to authenticate the client based on the features extracted by placing the retinotopic grid over his left eye, right eye or mouth.
The database used in our experiments is a subset of the first multi-modal verification oriented database collected by the M2VTS consortium. It consists of four series of images of 37 people taken at different periods of time. Each series contains two to four frontal images of each subject. Appearance of subjects varies widely across the different series. Tests are performed according to a leave-one-out rotation scheme knows as "Brussels Protocol". At each step, one person is removed from the database to act as impostor. Also, one entire image series is set aside for client tests. The 36 persons remaining act as registered clients, and their three image series are available for training. By choosing the impostor and the test series in all possible ways, 36x37x4=5328 essentially different client tests and the same number of impostor tests can be generated.
Experiments performed by manually placing the retinotopic sampling grid over the eyes and mouth of the subjects in the M2VTS database indicate that an Equal Error Rate as low as 0.2% can be achieved. This constitutes a dramatic improvement over other techniques, which yielded a 5% EER on the same database.