Similarity verification of kinship pairs using metricized emphasis

Abstract

Kinship verification is the determination of the validity of biological ties or kinship between two or more individuals, giving insights about genetic trait inheritances and other applications like forensic investigations. This paper presents a deep learning approach to kinship verification that methodically evaluates the similarity between images of kin. The proposed approach, Age-Modified Metricized Filtering (AMMF), begins by augments images via a Cycle-Generative Adversarial Network setup for aging child images, which increases facial parameters and reduces age gap. It then quantifies genetic inheritance by a novel method, Metricized Weight-based Emphasis Filtering, which reconciles facial proportions between the older and younger generation, and then uses Siamese networks for feature embedding and similarity evaluation. The approach is evaluated on a merged dataset of KinFaceW-I and KinFaceW-II, and achieves state-of-the-art performance. The results are suitable for real-world applications, achieving a training accuracy, AUC, and contrastive loss of 97.4%, 0.74 and 0.11 respectively. The approach also achieves 89.41% and 87.86% training accuracy on FIW and TSKinFace datasets respectively. This will contribute toward an accurate determination of the validity of kinship ties, thus contributing to tasks like image management, genealogical research, and criminal investigations.