Can we do better in Unimodal Biometric Systems? A Novel Rank-based Score Normalization Framework for Multi-sample Galleries Panagiotis Moutafis and Ioannis A. Kakadiaris Computational Biomedicine Lab, Department of Computer Science University of Houston, 4800 Calhoun Rd. Houston, TX 77004 {pmoutafis,ioannisk}@uh.edu Abstract The large amount of research on multimodal systems raises an important question: can we extract additional in- formation from unimodal systems? In this paper, we pro- pose a rank-based score normalization framework that ad- dresses this problem when multi-sample galleries are avail- able. The main idea is to partition the matching scores into subsets and normalize each subset independently. In ad- dition, we present two versions of our framework that: (i) use gallery-based information (i.e., gallery versus gallery scores), and (ii) update available information in an online fashion. We use the theory of Stochastic Dominance to il- lustrate that the proposed framework can increase the sys- tem’s performance. Our approach: (i) does not require tun- ing of any parameters, (ii) can be used in conjunction with any score normalization technique and any integration rule, and (iii) extends the use of W-score normalization to multi- sample galleries. While our approach is better suited for an Open-set Identification task, we also demonstrate that it can be used for a Verification task. In order to assess the performance of the proposed framework we conduct ex- periments using the BDCP Face database. Our approach improves the Detection and Identification Rate by 14.87% for Z-score and by 4.82% for W-score. 1. Introduction In this paper, we focus on the Open-set Identification task for unimodal systems when multiple biometric sam- ples per gallery subject are available. The Open-set Identi- fication is a two-step process: (i) determine whether a probe is part of the gallery, and (ii) return the corresponding iden- tity. The most common approach is to select the maximum matching score for a given probe and compare it against a given threshold. In other words, we match the probe against the gallery sample that appears to be the most similar to it. As a result, the Open-set Identification problem can be con- sidered to be a hard Verification problem. A detailed discus- sion is presented by Fortuna et al. [6]. This is not the only reason why the Open-set Identification task is a hard prob- lem. Each time that a subject submits its biometric sample to the system there are a number of variations that may oc- cur (e.g., differences in pose, illumination and other condi- tions during data acquisition). Consequently, each time that a different probe is compared against the gallery, the match- ing scores obtained follow a different distribution. One of the most efficient ways to address this problem is score nor- malization. Such techniques map scores to a common do- main where they are directly comparable. As a result, a global threshold may be found and adjusted to the desired value. Score normalization techniques are also very use- ful when combining scores in multimodal systems. Specif- ically, different classifiers from different modalities pro- duce heterogeneous scores. Normalizing these scores be- fore combining them is thus crucial for the performance of the system [8]. Even though this paper is not focusing on multimodal systems, the relative results are very useful in understanding the intuition behind the proposed approach. For the rest of this paper, we consider the following sce- narios: (i) the gallery set is comprised of multiple samples per subject from a single modality, and (ii) the gallery set is comprised of a single sample per subject from different modalities. We shall refer to the former scenario as uni- modal and to the latter as multimodal. We note that the in- tegration of scores in the unimodal scenario is an instance of the more general problem of combining scores in the mul- timodal scenario [12]. To distinguish between the two we say that we integrate scores for the former while we com- bine scores for the latter. We notice that a search in Google Scholar for the last ten years (i.e., 2002-2012) returns 322 papers that include the terms multimodal and biometric in their title while only eight entries are found for the unimodal case. The question that arises is whether there is space for improvement in the performance of unimodal systems. In this paper, we propose a rank-based score normal- ization framework that is suitable for unimodal systems __________________________________________________ ICB-2013, 6th International Conference on Biometrics ________________________________________________________ _________________________________________________ ICB-2013 June 4-7, 2013 Madrid, Spain