Synthetic Biometrics with Life-Like Appearance

S has a unique requirement to evaluate the performance of biometrics in extremely large systems. To address the requirement, we propose a novel and promising approach including three subsystems to generate large-scale realistic synthetic iris, fingerprint and face images matching the statistical and performance attributes of real-world data. For iris synthesis, we propose a novel and efficient solution, which uses patch-based sampling techniques. For fingerprint synthesis, we use Gabor-like space-variant filters for iteratively expanding an initially empty image containing just one or a few seeds. For face synthesis, we propose to develop the software on top of the commercial face generation software to generate a large-scale facial image dataset. To generate synthetic biometrics with life-like appearance, we will utilize an extensive set of biometrics features to transform synthetic biometrics into biometrics images which exhibit life-like characteristics. After generating independent biometric dataset, analysis tools will be developed to ensure all synthetic images have similar visual appearance, and texture/quality distribution compared with the real instances. After texture/quality and performance analysis, the qualified biometric images will be combined to generate a large-scale multimodal biometric dataset.