A multiple datasets deep learning approach for kinship recognition from ear images

In kinship recognition tasks, the face has traditionally been the primary biometric modality; in contrast, kinship recognition using ear images remains a largely unexplored area. However, ear width remains relatively constant throughout adulthood. Therefore, the ear can be considered a more stable biometric modality than the face. In our method, we used a Siamese network with convolutional sub-networks to extract the features from the ear images (VGG-16, ResNet-50, ResNet-152, and ConvNeXt-T). The extracted features are subsequently passed to a Siamese network classifier, which determines whether a kinship relation exists between the two input images. In our work, we used multiple datasets, including KinEar, Meng, EarKinshipVN, as well as modifications of the hyperparameters. Our results outperformed the method of Dvoršak et al., with the leading ResNet-152 and ConvNeXt-T as convolutional sub-networks.