Entropy-based template analysis in face biometric identification systems

The accuracy of a biometric matching algorithm relies on its ability to better separate score distributions for genuine and impostor subjects. However, capture conditions (e.g. illumination or acquisition devices) as well as factors related to the subject at hand (e.g. pose or occlusions) may even take a generally accurate algorithm to provide incorrect answers. Techniques for face classification are still too sensitive to image distortion, and this limit hinders their use in large-scale commercial applications, which are typically run in uncontrolled settings. This paper will join the notion of quality with the further interesting concept of representativeness of a biometric sample, taking into account the case of more samples per subject. Though being of excellent quality, the gallery samples belonging to a certain subject might be very (too much) similar among them, so that even a moderately different sample of the same subject in input will cause an error. This seems to indicate that quality measures alone are not able to guarantee good performances. In practice, a subject gallery should include a sufficient amount of possible variations, in order to allow correct recognition in different situations. We call this gallery feature representativeness. A significant feature to consider together with quality is the sufficient representativeness of (each) subject's gallery. A strategy to address this problem is to investigate the role of the entropy, which is computed over a set of samples of a same subject. The paper will present a number of applications of such a measure in handling the galleries of the different users who are registered in a system. The resulting criteria might also guide template updating, to assure gallery representativeness over time.