Multilinear subspace learning using handcrafted and deep features for face kinship verification in the wild

In this paper, we propose a new multilinear and multiview subspace learning method called Tensor Cross-view Quadratic Discriminant Analysis for face kinship verification in the wild. Most of the existing multilinear subspace learning methods straightforwardly focus on learning a single set of projection matrices, making it difficult to separate different classes. To address this issue, the proposed approach mutually learns multi-view representations for multidimensional cross-view matching. In order to decrease the effect of the within class variations for each mode of the tensor data, the proposed approach integrates the Within Class Covariance Normalization. Moreover, we propose a new tensor face descriptor based on the Gabor wavelets. Besides, we investigate the complementarity of handcrafted and deep face tensor features via their fusion at score level using the Logistic Regression method. Our extensive experiments demonstrate that the proposed kinship verification framework outperforms the state of the art, achieving 95.14%, 91.83% and 93.58% verification accuracies on Cornell KinFace, UB KinFace and TSKinFace face kinship datasets, respectively.

Automated summary

The proposed Tensor Cross-view Quadratic Discriminant Analysis method addresses the challenge of separating different classes in face kinship verification by learning multi-view representations. By integrating Within Class Covariance Normalization, the approach effectively decreases the impact of within class variations. Through fusion of handcrafted and deep face tensor features at score level, using Logistic Regression, the proposed method outperforms existing methods in verification accuracies.