Features extraction for low-power face verification

Mobile communication devices now available on the market, such as so-called smartphones, are far more advanced than the first cellular phones that became very popular one decade ago. In addition to their historical purpose, namely enabling wireless vocal communications to be established nearly everywhere, they now provide most of the functionalities offered by computers. As such, they hold an ever-increasing amount of personal information and confidential data. However, the authentication method employed to prevent unauthorized access to the device is still based on the same PIN code mechanism, which is often set to an easy-to-guess combination of digits, or even altogether disabled. Stronger security can be achieved by resorting to biometrics, which verifies the identity of a person based on intrinsic physical or behavioral characteristics. Since most mobile phones are now equipped with an image sensor to provide digital camera functionality, biometric authentication based on the face modality is very interesting as it does not require a dedicated sensor, unlike e.g. fingerprint verification. Its perceived intrusiveness is furthermore very low, and it is generally well accepted by users. The deployment of face verification on mobile devices however requires overcoming two major challenges, which are the main issues addressed in this PhD thesis. Firstly, images acquired by a handheld device in an uncontrolled environment exhibit strong variations in illumination conditions. The extracted features on which biometric identification is based must therefore be robust to such perturbations. Secondly, the amount of energy available on battery-powered mobile devices is tightly constrained, calling for algorithms with low computational complexity, and for highly optimized implementations. So as to reduce the dependency on the illumination conditions, a low-complexity normalization technique for features extraction based on mathematical morphology is introduced in this thesis, and evaluated in conjunction with the Elastic Graph Matching (EGM) algorithm. Robustness to other perturbations, such as occlusions or geometric transformations, is also assessed and several improvements are proposed. In order to minimize the power consumption, the hardware architecture of a coprocessor dedicated to features extraction is proposed and described in VHDL. This component is designed to be integrated into a System-on-Chip (SoC) implementing the complete face verification process, including image acquisition, thereby enabling biometric face authentication to be performed entirely on the mobile device. Comparison of the proposed solution with state-of-the-art academic results and recently disclosed commercial products shows that the chosen approach is indeed much more efficient energy-wise.