Deep discriminant generation-shared feature learning for image-based kinship verification

Abstract

Kinship verification based on facial images in the wild is an interesting and challenging research topic in the fields of computer vision. It has many potential applications, e.g., missing children search, image annotation, etc. In practice, the biggest obstacle in kinship verification is that there usually exists large divergence between the images of parent and children. How to effectively tackle such challenge is important for improving kinship verification performance. To this end, we proposed a novel Deep Discriminant Generation-Shared Feature Learning for Image-Based Kinship Verification (D2GFL) method for kinship verification, which consists of a two-stream generation-specific feature learning module and a generation-shared discrepancy reducing module. The generation-specific feature learning module can learn parent-specific and child-specific features. The generation-shared module aims to reduce the divergence between facial images of parent and child. In order to further relieve the cross-generation difference and improve the discriminability of learned features, we also design a cross-generation difference loss term and an intra-generation discriminant loss term, which simultaneously make use of the local and holistic features, as well as the family identity information. Experimental results on several widely used kinship datasets are presented to validate the effectiveness of our proposed approach by comparing with the state-of-the-art kinship verification methods. Specially, our approach can improve the average verification accuracy at least by 3.5%, 1.5% and 0.8% on KinFaceW-I, Cornell, and TSKinFace datasets, respectively.