A deep learning algorithm to address kinship verification integrating age transformation techniques applied to the family images and model tuning methodologies

Abstract

This thesis addresses the challenge of verifying familial relationships through facial features, which is often complicated by age-related variations. Traditional kinship verification models struggle to account for these changes, resulting in decreased accuracy. Accurate kinship verification is crucial for various applications, including forensic investigations, family reunion efforts, and social media analysis to mention some implementations of kinship verification. However, traditional kinship verification models struggle to account for these changes, resulting in decreased accuracy. To tackle this, the objective was to enhance kinship verification by integrating age transformation techniques into a deep learning model. The proposed solution involved applying the LATS (Learnable Age Transformation Synthesis) algorithm to transform images into different age ranges, making familial traits more recognizable. A deep learning model using a Siamese network architecture was trained on the Families in the Wild (FIW) dataset, with age transformations applied at 5, 15, and 30 years to address the model’s ability to identify kinship relationships of mother, father and children. The model was evaluated using accuracy, F1-score, and Mean Squared Error (MSE) across different transformation scenarios. The results showed an overall accuracy of 87 %, with the best performance in father-children relationships at a 5-year transformation and in mother-children relationships at a 15-year transformation, demonstrating the model’s effectiveness in capturing age-specific familial traits.