Kinship verification through facial images is important due to its usage in many significant and ethically sensitive applications such as forensics, humanitarian efforts, and identity verification. Although many deep learning models were employed to verify the kinship between people, the accuracy of these models needs enhancement especially in age and gender variances. As well as missing transparency in decision-making process due to “black box” nature of these architectures restricts user trust. In this study, we propose EnhancedKin-shipNet, a novel deep learning architecture that combines a ResNet50 backbone with an attention-based feature fusion mechanism and XAI integration to accurately classify parent–child facial relationships, improve the performance in comparison with state-of-the-art and boost the user trust as well as transparency in decision making. Two benchmark datasets, KinFaceW-I and KinFaceW-II, are analyzed in this study. KinFaceW-I contains 533 images while KinFaceW-II includes 1000 images. Both datasets consist of four classes of kinship relationships: Mother-Daughter (M-D), Father-Daughter (F-D), Mother-Son (M-S), and Father-Son (F-S). We evaluate the model three times, the first and second through using KinFaceW-I and KinFaceW-II individually. While the third evaluation is implemented by combination of both datasets. Individual KinFaceW-I achieved 94.44% accuracy while individual KinFaceW-II achieved 96%. The accuracy achieved by both datasets is 93.46%. We employ Grad-CAM and LIME to interpret model decisions through indicating the regions that mainly affect verification. Our results provide end-to-end and transparent framework with competitive performance, facilitating accurate, trustworthy and understandable verification of kinship relationships.
Al-Falluji et al. (Sun,) studied this question.