Kinship Measurement on Face Images by Structured Similarity Fusion

Kinship verification via facial images is an emerging problem in computer vision and biometrics. Recent research has shown that learning a kin similarity measurement plays a critical role in constructing a vision-based kinship measurement system. We propose in this paper a new similarity metric learning method for kinship measurement on human faces. To this end, we first extract multiple feature representations for each face image using different face descriptors. Then, multiple sparse bilinear similarity models (one for each view) are jointly learned by using joint structured sparsity-inducing norms, such that the similarity score of a pair of child-parent images is consistently higher than those of the pairs without kinship relations while leveraging the interactions and correlations among the multiview data to obtain the fused and higher level information. We also derive an effcient algorithm to solve the formulated nonsmooth objective. Experimental results on kinship data sets show that our method achieves competitive or better accuracy performance in comparison with the state-of-the-art multimetric learning-based kinship verification methods but enjoys the superiority in computational efficiency, making it more practical for vision-based kin measurement applications.