Kinship Detection Based on Hand Geometry Using ResNet50 Model for Feature Extraction

Kinship (family relationships) detection is important in many domains; it can be used in forensic investigations, adoption, biometric security, and more. It is particularly necessary in times of conflict and natural disasters, such as earthquakes, as it can help with reunions, and searches for missing persons. The most popular and very accurate means of establishing kinship is DNA analysis. Another method which is non-invasive is kinship estimate using facial images and computer vision accompanied with machine learning algorithms. Every component of the human body contains embeddedinformation that may be taken out and used for that persons identification, verification, or classification. Finding characteristics that every family has in common is the foundation of kinship detection. This paper examines a novel approach of kinship detection using the hand geometry. Deep transfer learning using the ResNet50 model was used to extract geometrical features from hand images. A neural network classifier was designed and trained to predict kinship and assembled as a top layer for the ResNet model. The test accuracy of this novel methodology was 92.8% yielding to the hand has geometrical features that can be used to detect kinship, and that the proposed method is a possible potential way to identify kinship. We built our own hand image dataset that contains kinship ground truth metadata since there were no such datasets before. We called it Mosul Kinship Hand (MKH) dataset, which includes 648 photos of 81 people from 14 households (8 different hand images per person), and it was used in this research.