Optimization based approaches for score and rank level fusion of multimodal biometrics

Abstract

Abstract newlineUnimodal biometric systems have several limitations, like inter-class newlinesimilarity, non-universality, and susceptibility to circumvention. Multiple newlinebiometric modalities are fused to overcome these issues. Here, newlinethe fusion is mainly applied to the information from multiple biometric newlinemodalities. Fusion in multimodal biometrics is performed at newlinevarious levels, such as sensor level, feature level, score level, rank level newlineand decision level. The score and rank level fusion are the two widely newlineapplied fusion techniques for multimodal biometrics. In the context of newlinean identification task, these methods fuse matching score lists or rank newlinelists from different biometric modalities into a single score or rank list, newlinerespectively. newlineIn this thesis, rank and score level fusion problems are formulated as newlineoptimization problems. Here, the objective is to find a fused list (for newlineeither rank or score). The fused list minimizes a weighted summation newlineof distances of the fused list with the input lists derived from individual newlinebiometric modalities. The stated distance between a pair of input newlinelists is computed using the weighted Spearman footrule distance metric. newlineGenetic algorithm based and particle swarm optimization based newlinefusion approaches (at rank level and at score level) are proposed in this newlinethesis to solve the stated optimization problems. For initial work, each newlinemodality is assigned equal significance (weight). Furthermore, the newlinequality-based weight estimation approach is presented in this work newlineto enhance the performance of proposed optimization based fusion newlineapproaches. The quality-incorporated optimization based fusion approaches newlineperform better than the equal weight based optimization newlineapproaches. newlineThe adopted optimization based fusion approaches (genetic algorithm newlineand particle swarm optimization) are meta-heuristic algorithms. newlineThese algorithms iteratively search for the optimal solution in a large newlinevi newlinesearch space. Therefore, these approaches take immense number of newlineiterations to reach to the optimal solution. An approach to reduce newlineth