Image-based Kinship Verification Using Dual VGG-Face Classifier

In this paper, we investigated dual VGG-Face convolutional neural network classifier for image-based kinship verification problems. The proposed classifier is formed by paralleling convolutional layers of VGG CNN architecture and combined it with several fully-connected layers. Although the VGG CNN architecture is consists of huge parameters, the number of parameters in the classifier will be reduced by more than 80% by removing the fully-connected layers of the original classifier. We use the multi-task loss function for the training process to ensure that the features learned by the classifier are also can be used for family classification. Experiments on the FIW kinship verification dataset show that dual VGG-Face CNN classifiers can achieve an average accuracy of 64.71% on a single classifier and 65.49% on ensemble configuration. Based on our experiments, the lowest accuracy is produced for second-generation kinship (grandchild - grandparent) which have a low number of examples on FIW kinship.