Abstract—Facial appearance affects how humans interact with each other. It is also how relatives are visually identified to determine how social interactions proceed. Humans can identify kin relations based only on the face. Intrinsically, giving the ability to detect kin relations to computers can improve their usefulness in our daily lives. This research proposes a solution to the kinship verification problem with a novel non-context-aware approach. The approach is validated using a dataset with large age variation upon which is applied the proposed Deep Linear Metric Learning (DLML). The method leverages multiple deep learning architectures trained with massive facial datasets. The knowledge acquired on traditional facial recognition tasks is re-purposed to feed a linear metric learning model. The proposed method was able to achieve better performance than other context-aware methods on tests that are inherently more difficult than ones used on previous methods with the UB Kinface dataset. The results show that the method can use the knowledge of deep learning architectures trained to perform mainstream facial recognition tasks with massive datasets to solve kinship verification on the UB Kinface database with robustness towards large age differences present on the dataset. The method also offers enhanced applicability when compared to previous methods in real-world situations, because it removes the necessity of knowing/detecting and treating large age variations to perform kinship verification. Additional tests were also performed at the KinFaceW-I and KinFaceW-II datasets to further assess performance of the method.

Index Terms—kinship verification, deep learning, feature re-purposing, metric learning, facial recognition, convolutional neural network

Cite: Diego Lelis and Dibio L. Borges, "Facial Kinship Verification with Large Age Variation Using Deep Linear Metric Learning," Journal of Image and Graphics, Vol. 7, No. 2, pp. 50-58, June 2019. doi: 10.18178/joig.7.2.50-58