Yüz biyometrik doğrulama sistemleri için kızılötesi aydınlatma altında gölge özelliklerinin araştırılması

With the advancements in printing technologies and high-resolution screens, planar presentation attacks are becoming hard to detect with current methodologies. 3D facial structure extraction seems to be a promising way to go in Liveness detection. In this thesis, a simple approach has been proposed that benefits from the shading of 3D facial structure. CNN models were trained to estimate the lightning direction on the face by using a synthetic dataset generated for this purpose with a Unity application developed: unityStudio. A dual IR LED setup is used to suppress environmental lighting and to create a dominant lightning direction. The thesis is, by comparing the model response with prior knowledge of lightning direction, the presence of a 3D face structure will be able to be confirmed. Experiments were made on state-of-the-art CNN models and simple CNN implementations. Successful CNN models are trained on synthetic data and their real-world performances were criticized.