Super-Sparse Biometric Recognition

Many modern biometric recognition systems require using a small set of reference templates per client to save memory and computational resources during client verification and identification.

However, both the reference templates and the combination of the corresponding matching scores are often heuristically chosen. 

Under this scenario, we are investigating a well-grounded approach, called super-sparse biometric recognition, capable of outperforming state-of-the-art methods both in terms of recognition accuracy and number of required reference templates, by jointly learning an optimal combination of matching scores and the corresponding subset of templates.

For each client, it consists of learning a support vector machine (SVM) using the given matching algorithm as the kernel function. This automatically determines a set of reference templates, that we call support biometrics, corresponding to the SVM’s support vectors. It then drastically reduces the number of templates, without affecting recognition accuracy, by learning a set of virtual biometric traits as well-principled transformations of the initial support biometrics (namely, the initial support vectors).

Fig. 1. A conceptual example of super-sparse biometric recognition. Assume biometric traits (e.g., face images) are given as samples in a vector space, and red and blue points respectively denote genuine and impostor claims for a given client. In the left plot, the black line represents the SVM's decision function obtained by learning an SVM on the given data, yield 12 support vectors (i.e., support faces), highlighted with black circles. In the right plot, we report the reduced solution obtained with our approach. Despite being very similar to the one in the left plot, it only requires 4 virtual support templates (faces), reducing the complexity (i.e., number of matchings) required for client verification of three times.


The use of a very small set of support biometric templates makes the decisions of our approach also easily interpretable for designers and end users of the recognition system.

Fig. 2. An example of virtual faces found by our approach for a given client. The first virtual face well resembles the genuine client, while the remaining three faces are combination of impostors. These latter images, as in cohort-based biometric verification, help improving the correct verification of the genuine client.


We have implemented super-sparse biometric recognition for face verification (see our paper “Sparse Support Faces”), and we are now working on its application to fingeprints and identification problems.