The aim of this paper is to investigate several number of normalization method and fusion rule at matching score level. Fusion at feature level can increase the discrimination power in the feature space by generating high dimensional fuse feature vector. However this approach produces high computational cost. Fusion at matching score level utilizes the matching output from different matching module to form a single value for the decision process. A normalization method is utilized to preprocess the matching score before a fusion process. The analysis shows sum rule and min-max normalization produce the best recognition rates 98% and verification rates GAR = 94% at FRR = 0.1%.
