Modified Siamese Convolutional Neural Network for Fusion Multimodal Biometrics at Feature Level
Husam Imad Abdulrazzaq, Nidaa Flaih Hassan · 2019
Feature level fusion in multimodal biometric systems is a significant topic in personal identification researches. This fusion technique faces two main problems: incompatible features of dissimilar modalities, and the curse of dimensionality. This paper presents a solution for these problems, by proposing a modified Siamese Convolutional Neural Network (Siamese-CNN) structure, which consists of two identical sub-networks with a shared set of weights. These sub-networks would be joined together and pass through a single, fully connected layer. The proposed structure is implemented as a multimodal biometric identification system for fusion face and palm veins modalities at the feature level. Each modality images pass through its corresponding sub-network. The proposed structure is responsible for feature extraction, feature fusion, and classification steps. The structure is experimented on a simulated database for 50 persons are collected from two well-known public databases; six experiments are performed to demonstrate the efficiency and consistency of the structure. The results prove that the proposed structure overcomes the problems of feature level fusion technique by increasing the consistency between the two dissimilar modalities and high identification speed. In addition to the structure is robust against imposters and achieves high accuracy, which is 99.33% ± 0.67.