SECCION III: COMPORTAMIENTOS DE RIESGO EN LA VIDA SEXUAL Y REPRODUCTIVA
N° EN RELACIÓN A LA PRIMERA VEZ DE ACTIVIDAD SEXUAL
This section discusses the levels of fusion at which multi-modal biometric systems may typically operate. During the discussion, the relationship between the formal levels of data fusion covered in section 2.3.1.2 and those used in a multi-modal biometric system will be noted.
As discussed in section 2.2.2 (and illustrated in Figure 2.1), a generic uni-modal biometric system consists of four modules (Ross and Jain, 2004; Faundez-Zanuy, 2009). For a generic multi-modal biometric system, a fusion module must be added (refer Figures 2.6, 2.7, and 2.8).
A description of the roles for the four modules in a generic uni-modal biometric system was provided in section 2.2.2. The fusion module (added to the generic multi-modal biometric system) is responsible for combining the data from multiple sources, and the proposed level of fusion determines its location in the system.
Figure 2.5 demonstrates that fusion can occur before the matching process or after the matching process (Poh and Kittler, 2008). If fusion occurs before the matching process, data may be fused at either the sensor or feature level. If fusion occurs after the matching process, data may be fused at either the condence score, rank, or decision levels.
Figure 2.5: Data Fusion Levels In Multi-Modal Biometrics The image was sourced from Poh and Kittler, (2008).
The data fusion levels in a multi-modal biometric system, are described as follows (Ross and Govindarajan, 2005; Poh and Kittler, 2008):
• Sensor Level. Raw data represents the richest source of information (though it may possibly be contaminated by noise). Raw data can be processed such that a new, single vector, consisting of the integrated raw data, is obtained. The newly created raw data vector may be processed directly or features may be extracted from it. An important caveat regarding sensor level fusion is that it is only possible to fuse data when samples of the same biometric trait are used (Nandakumar, 2008). That is, where raw data from multiple instances using the same sensor, or samples from multiple sensors, provide readings of the same biometric. Consequently, this level of fusion is rarely attempted in multi-modal biometrics because raw data (from multiple modes) can not be meaningfully combined.
Sensor level fusion in multi-modal biometrics relates to the rst level of the formal data fusion levels covered in section 2.3.1.2.
• Feature Level. This fusion level uses data collected from the feature extraction process (Osadciw et al., 2003). Researchers believe that feature level fusion will result in accurate and robust authentication, because data at this level
is closer to raw datathan the subsequent fusion levelsand maintains more discriminatory information than those levels (Ross and Jain, 2004).
However, in order to achieve a high level of system performance, extensive pro- cessing is typically required (refer section 2.3.2.2). Feature extraction typically requires the selection of salient features, from the independent data sources, that best represent the entity and can provide recognition accuracy (Poh and Kittler, 2008).
Figure 2.6 illustrates the process ow recommended for feature level fusion. Firstly, the biometric capture modules are responsible for acquisition of the biometric characteristics. In Figure 2.6 there are two such capture modules, indicating a dual mode biometric system.
Following acquisition, is the feature extraction module (for each mode). At this stage, the features relating to each individual mode remain separate (that is, they remain separate feature vectors). Data alignmentto bring the fea- tures from multiple independent sources into alignment (as discussed in section 2.3.1.3)will usually be required prior to fusion. Feature selection may also be requiredprior to fusionto reduce the size of the nal feature set; otherwise the fused feature set may suer from the `curse of dimensionality' (Ross and Govindarajan, 2005).
After feature extraction and data alignment (and possibly feature selection), the fusion module combines the feature vectors corresponding to each inde- pendent source. If data alignment was necessary, and had been successfully performed, the fusion process becomes relatively simple. In some literature re- lated to feature level fusion, the feature vectors are simply concatenated (Ross et al., 2001; Ross and Jain, 2004; Faundez-Zanuy, 2009; Nandakumar, 2008). The fused vector of features is then passed to the matching module, where it is compared to a registered template. The template feature vector must have been processed in the same manner, and result in the same format, as the query feature vector. The output from the matching module is then passed to the decision module, where the nal classication decision is made.
It should be apparent that the complex and intensive processing required for feature level fusion (to avail the system of the advantages of data fused at this level) is mainly associated with feature pre-processing for the appropriate representation of features, feature selection, and data alignment.
Figure 2.6: Feature Lev el Data Fusion The image w as sourced from Maltoni et al., (2003).
Feature level fusion in multi-modal biometrics relates to the second level of the formal data fusion levels covered in section 2.3.1.2.
• Condence (or Matching) Score Level. Fusion at this level occurs after a matching module (for each mode) has determined a condence or matching score (Osadciw et al., 2003; Indovina et al., 2003). The condence score is a measure of similarity between the query and registered biometric feature vec- tors (Nandakumar, 2008). This level of fusion is the most commonly employed and researched of all fusion levels in multi-modal biometrics (Poh and Kittler, 2008; Nandakumar, 2008).
Figure 2.7 illustrates the process ow recommended for condence score level fusion. The biometric capture and feature extraction modules perform their tasks as they would for feature level fusion. However, fusion at the con- dence score level does not involve the integration of the feature vectors. The individual feature vectors are passed to the separate matching modules for comparison with registered templates. That is, instead of a one vector com- parison (as with feature level fusion), there are multiple vectors compared with their corresponding templates (in the appropriate matching modules).
A condence score is calculated in each matching module, according to the above comparison, and passed to the fusion module. The condence score is essentially a probability score in the continuous domain (typically in the interval[0,1]). The scores from the individual matching modules are combined
in the fusion module, into a single scalar value (typically in the intervals [0,1]
or[0,100]), which is then passed to the decision module. The decision module
makes the nal classication decision, based on this fused condence score. Fusion at this level requires less processing (than feature level fusion) in order to achieve an appropriate level of system performance.
Condence score level fusion in multi-modal biometrics is not specically de- ned among the formal data fusion levels covered in section 2.3.1.2. One could conjecture that it may t into the decision level of the formal data fusion levels.
Figure 2.7: Cond ence Score Lev el Data Fusion The image w as sourced from Maltoni et al., (2003).
• Rank Level. Fusion at this level is more relevant to identication than verica- tion (Ross and Govindarajan, 2005). For every enrolled identity in a database, the condence scores for each mode are sorted in descending order and ranked such that the lowest rank indicates the `worst' match and the highest rank indi- cates the `best' match (with all other condence scores appropriately ranked). A fusion method is used to consolidate the individual condence scores. Rank level fusion in multi-modal biometrics is not specically dened among the formal data fusion levels covered in section 2.3.1.2, though being a subset of condence score level, it may t into the decision level of the formal data fusion levels.
• Decision Level. Fusion at this level is the most abstract, where accept/reject decisionsfrom multiple data sourcesare consolidated into one nal classi- cation decision (Osadciw et al., 2003).
Figure 2.8 illustrates the process ow recommended for decision level fusion. The biometric capture, feature extraction, and matching modules perform their tasks as they would for condence score level fusion. However fusion at the decision level, does not involve the integration of the condence scores. The individual condence scores are passed to the decision modules, where accept/reject decisions (output as boolean values) are made for each mode. Each decision module makes its classication decision (based on the condence score passed to it), and passes its decision to the fusion module. The multiple boolean values are integrated to generate the nal classication decision. Fusion at this level takes advantage of the processing performed by each mode's matching and decision modules to arrive at an individual decision. This makes decision level fusion more scalable than the other levels.
Fusion at this level requires the least amount of processing (compared to the other levels) in order to achieve an appropriate level of system performance. Decision level fusion in multi-modal biometrics relates to the third level of the formal data fusion levels covered in section 2.3.1.2.
Figure 2.8: Decision Lev el Data Fusion The image w as sourced from Maltoni et al., (2003).
The next section presents a review of some experimental eorts in multi-modal biometric authentication.