Abstract

Kinship verification from faces is a challenging task that is attracting an increasing attention in the recent years. The proposed methods so far are not robust enough to predict the kin between persons via facial appearance only. The initial studies using deep convolutional neural networks (CNN) have not shown their full potential as well, mainly due to limited training data. To mitigate this problem, we propose a new approach to kinship verification based on color features and extreme learning machines (ELM). While ELM aims to deal with small size training sets, color features are proven to provide significant enhancement over gray-scale counterparts. We evaluate our proposed method on three benchmark and publicly available kinship databases, namely KinFaceW-I, KinFaceW-II and TSKinFace. The obtained results compares favorably against some state-of-the-art methods including those based on deep learning.