With the security requirement increasing in personal authentication area,multimodal biometric recognition has become a hot research topic.Multimodal biometric fusion can be implemented at four levels: pixel level,feature level,score level,and decision level.Since the features contain higher discrimination information,feature level fusion is widely regarded as the most effective operation for improving the discrimination of multimodal finger features.However,feature level fusion may have some problems in practice,such as feature space incompatibility and curse of dimensionality.Therefore,a hypersphere granular fusion method based on fingerprint(FP),finger-vein(FV)and finger-knuckle-print(FKP)is proposed in this thesis.The main works are summarized as follow:First,LTT(Local Ternary Texton)and HOG(Histograms of Oriented Gradients)are applied to jointly describe gray distributions and texture features of finger images.Second,the feature vector of each modality is expressed as an atomic hypersphere granule in high-dimensional feature space.Then,a triangle is constructed based on the relative position relations of the atomic granules corresponding to three modalities.To take full advantage of the discrimination of an optimal modality,a fused hypersphere granule is generated by connecting centers between FV granule and the triangle inscribed circle.In this way,a new FV-biased triangle fusion model is established accordingly.Finally,a modified fuzzy inclusion measure is used to compute the similarity between two fusion hypersphere granules for feature matching.The experimental results show that the proposed hypersphere granular fusion method can integrate the features of FP,FV and FKP effectively,and achieve higher recognition accuracy and efficiency compared with the traditional fusion methods. |