ABSTRACT
The face-identification task is an important application of biometrics, which becomes extremely challenging in cases such as distinguishing between identical twins. Most approaches to overcome this problem involve using algorithms for facial local-feature extraction and texture analysis. In this chapter, rather than using hand-crafted features such as local features and developing algorithms based on texture analysis, a deep discriminative model based on the similarity metric and using global features is proposed. We use a similarity learning metric based on the Siamese architecture to leverage a large dataset for distinguishing between identical twins and look-alike persons. This chapter presents a proposed architecture and method for verification, and shows the ability of the deep convolutional neural networks (CNNs) in the challenging task of distinguishing between identical twins and includes the following: Siamese architecture has been applied as a discriminative model to map the input space to a target space in which the similarity metric is a simple euclidean distance. Transfer learning by fine-tuning on the targeted dataset has been demonstrated to be effective to increase the evaluation performance. Finally, the generalizability of the method has been verified by evaluation of the model on different datasets.
