Gender-Invariant Face Representation Learning and Data Augmentation for Kinship Verification

Different from conventional face recognition, the gender discrepancy between parent and child is an inevitable issue for kinship verification. Father and daughter, or mother and son, may have different facial features due to gender differences, which renders kinship verification difficult. In view of this, this paper proposes a gender-invariant feature extraction and image-to-image translation network (Gender-FEIT) that learns a gender invariant face representation and produces the transgendered images simultaneously. In Gender-FEIT, the male (female) face is first projected to a feature representation through an encoder, then the representation is transformed into a female (male) face through the specific generator. A gender discriminator is imposed on the encoder, forcing to learn a gender invariant representation in an adversarial way. This representation preserves the high-level personal information of the input face but removes gender information, which is applicable to cross-gender kinship verification. Moreover, the competition between generators and image discriminators encourages to generate realistic-looking faces that can enlarge kinship datasets. This novel data augmentation method significantly improves the performance of kinship verification. Experimental results demonstrate the effectiveness of our method on two most widely used kinship databases.