Robust Kinship Verification Using StyleGAN-Based Age–Gender Normalization and Siamese Metric Learning

Abstract

Facial kinship verification is a one-to-one classification problem. The objective is to determine whether a pair of facial images shares a blood relation. Due to its real-world applications, such as in social media analysis and the search for missing children, this problem has received greater attention in the fields of image processing and computer vision. However, age and gender differences between facial images of parents and children make it difficult to extract discriminative and similar features. Our research addresses this problem by normalizing the faces with respect to age and gender using e4e encoder and StyleGAN. In the first stage, we generate the parent and child images across age bands (10-15, 25-30, 55-60 years) and different genders. In the second stage, a Siamese network is used to extract and match facial features. We evaluated our model on two prominent datasets, KinFaceW-I and KinFaceW-II, containing kin faces in an unconstrained environment. In addition, we also proposed a newly constructed Indian kinship dataset KiN_IND, which contains face images from Indian demography, and present a comprehensive evaluation of proposed and state-of-the-art approaches on this dataset. Our proposed method achieves an average accuracy of 96.86% and 97.50% on KinFaceW-I and KinFaceW-II, respectively, outperforming existing state-of-the-art methods.