Leveraging implicit demographic information for face recognition using a multi-expert system

This paper describes a novel biometric architecture to implement unsupervised face recognition across varying demographics. The present proposal deals with ethnicity, gender and age, but the same strategy can be crafted for any mix of soft/hard biometrics, sensors, and/or methods. Our aim is not to explicitly distinguish demographic features of a subject (e.g., male vs. female). We rather aim at implicitly exploiting such information to improve the accuracy of subject identification. The role demographics plays in authentication has been reported by many recent studies. Exploiting demographic information can entail two possible strategies. Both require pre-determination of relevant demographic classes, that drive the choice of the best suited recognizer in a set of ad-hoc trained ones. In the first strategy, a human operator visually classifies demographic features of the subject to recognize, and runs the appropriate “strong” recognizer. In the second one, the identification of the most appropriate “strong” recognizer follows the results obtained from a set of upstream classifiers for soft biometrics. Both solutions are poorly suited to most real world applications, e.g., video - surveillance. Our architecture mediates recognition across different demographics without any pre-determination of demographic features. We still have different “strong” classifiers, each trained on a demographic class. The probe is submitted to all of them at once. A supervisor module estimates reliability of the single responses, and the most reliable result is returned. In this approach, classifier reliability is not a static feature, but it is estimated for each probe. The proposed multiple-expert system provides similar performance to pre-determination of demographics. Experimental results show higher flexibility, efficacy and interoperability. We also focus on interoperability across face datasets by adopting EGA (Ethnicity, Gender and Age) database as a benchmark, which is obtained by combining images from several publicly available face datasets.