Quasi Linear Sigmoid (score normalization)
The normalization module is very important in the context of multi-biometric systems, particularly when a parallel or hierarchical scheme is implemented. In general each subject in the output list of the first level subsystems (face and ear) is labelled with a numeric value in the range [0, ¥]. However, such value come from measurements performed on different features using different procedures; a direct combination of such values would give an incorrect result. According to this consideration it is clear the need to normalize the scores assigned by each biometry to individual subjects before combining them using any fusion rule. A number of different solutions have been proposed in literature to solve this problem. We researched a specific normalization function derived from the family of sigmoidal functions.
The function F(x) assures a pseudo-linear mapping for all values of x included in the interval [0, xmax]; if we admit some distortion, it also allows to normalize values of x greater than xmax, still guaranteeing the constraint F(x)<1. This is essential in those biometric systems where the value of xmax is not known in advance.
System Reliability Responses (system reliability)
There is an especially crucial aspect which is seldom considered, i.e. the definition of a measure for the response reliability of the single subsystems before their results are fused in an overall final response. The main reason can be found in the considerable difficulty implicit in binding response reliability to input data quality. The introduced mesure will be called System Response Reliability (SSR) and is based on the ability by a recognition method of separating genuine subjects from impostors in the most sharp and unequivocal way possible. The Figure shows two experiments in which red circles represents genuine users correctly acce