Multi-Level Fuzzy Score Fusion for Client Specific Linear Discriminant Analysis Based Face Authentication System

Abstract

A client specific linear discriminant analysis (CSLDA) based face authentication system has been developed with multi-level fuzzy score fusion. The CSLDA method provides two measures for authentication: distance to the client (Client Score) and distance to the mean of impostors (Impostor Score). A two-level multi-sample score fusion method has been proposed. A fuzzy inference module has also been developed to combine the scores of the CSLDA in the first level. The performance of fuzzy inference score fusion is then compared with several existing fusion methods and the conventional CSLDA face authentication system (without score fusion). Overall, the proposed fusion methods improve the performance of the algorithm and are more robust to variability of the inputs. Evaluation experiments were carried out with two different databases (AT&T and BANCA) where each contains face images of 40 subjects. Experimental results showed that the proposed multi-level fuzzy score fusion method improves the performance of the CSLDA based face authentication system compared to the other fusion techniques examined in this work.