Improving identification by pruning: A case study on face recognition and body soft biometric

Abstract

We investigate body soft biometrics capabilities to perform pruning of a hard biometrics database improving both retrieval speed and accuracy. Our pre-classification step based on anthropometric measures is elaborated on a large scale medical dataset to guarantee statistical meaning of the results, and tested in conjunction with a face recognition algorithm. Our assumptions are verified by testing our system on a chimera dataset. We clearly identify the trade off among pruning, accuracy, and mensuration error of an anthropomeasure based system. Even in the worst case of ±10% biased anthropometric measures, our approach improves the recognition accuracy guaranteeing that only half database has to be considered.