Exploiting Relative Entropy and Quality Analysis in Cumulative Partial Biometric Fusion

Abstract

Relative Entropy (RE) of individual’s biometric features is the amount of information that distinguishes the individual from a given population. This paper presents an analysis of RE measures for face biometric in relation to accuracy of face-based authentication, and proposes a RE-based partial face recognition scheme that fuses face regions according to their RE-ranks. We establish that different facial feature extraction techniques (FET) result in different RE values, and compare RE values in PCA features with those for a number of wavelet subband features at different levels of decomposition. We demonstrate that for each of the FETs there is a strong positive correlation between RE and authentication accuracy, and that increased image quality results in increased RE and increased authentication accuracy for all FETs. In fact, severe image quality degradation may result in more than 75% drop in RE values. We also present a regional version of these investigations in order to determine the facial regions that have more influence on accuracy and RE values, and propose a partial face recognition that fuses in a cumulative manner horizontal face regions according to their RE-ranks. We argue that the proposed approach is not only useful when parts of facial images are unavailable but also it outperforms the use of the full face images. Our experiments show that the required percentage of facial images for achieving the optimal performance of face recognition varies from just over 1% to 45% of the face image depending on image quality whereas authentication accuracy improves significantly especially for low quality face images.