Home // International Journal On Advances in Life Sciences, volume 14, numbers 3 and 4, 2022 // View article

Explainable Kinship: A Broader View on the Importance of Facial Features in Kinship Recognition

Authors:
Britt van Leeuwen
Arwin Gansekoele
Joris Pries
Etienne van de Bijl
Jan Klein

Keywords: kinship recognition; StyleGAN2; Families-in-the- Wild; feature importance; transfer learning

Abstract:
Kinship Recognition, the ability to distinguish be- tween close genetic kin and non-kin, could be of great help in society and safety matters. Previous studies on human kinship recognition found interesting insights when looking for the most important features. Results showed that analyzing only the top half of a face gives equal or even better performance compared to analyzing the whole face. In this paper, we aim to find the important features for automated kinship recognition based on the theory of human kinship recognition; this set of features was researched using features from pre-trained metrics from the StyleGAN2 model. Three different experiments were performed focusing on different aspects of facial features. We found that the most important facial features from the selection of 40 features are mostly focused on the facial hair traits. Furthermore, age-related features were found to be very important. This set of features does not entirely comply with the set of features important in human kinship recognition. Previous research has shown human kinship recognition performance does not decrease when removing the bottom half of the image of the face. In contrast, our results show that for automated kinship recognition, removing either the bottom or the top half of a face results in a decrease in the performance of our classifiers. Moreover, only using a selection of facial features corresponding with the important features in human kinship recognition did not prove to be sufficient for the task of Kinship Recognition.

Pages: 89 to 99

Copyright: Copyright (c) to authors, 2022. Used with permission.

Publication date: December 31, 2022

Published in: journal

ISSN: 1942-2660