Facial Soft Biometrics Detection on Low Power Devices

Soft biometric traits have been proven to enhance person identification accuracy, when used complementary to primary biometric traits. They present a series of advantages such as compliance to the human language, robustness to low quality data, non-intrusive and consent free acquisition, and privacy preservation, increasing their applicability in realistic conditions. They can be extracted from a variety of individual modalities, with the human face being considered as the most informative source of attributes, as it provides rich geometrical and texture features. Recent advances in computer vision have allowed the accurate detection of such features under varying, non-ideal capturing conditions, with this increase in detection capacity, however, coming at the cost of high computational complexity. Meanwhile, the research and market interest has shifted towards the implementation of such methods on low power devices (i.e mobile phones), with data security concerns favoring on-device offline computation instead of cloud-based services. Towards this end, and taking into consideration recent advances in computationally efficient CNN design and multitask learning, we propose a novel CNN architecture, suitable for real time implementation on low power devices, which simultaneously performs gender, age, race, eyes state, eyewear, smile, beard and moustache estimation from unconstrained face images. The architecture employs the Mobilenet architecture and exploits the correlation between the individual biometric features, performing comparably to three state-of-the-art face analysis systems, while requiring significantly lower computational resources.