التحقق من أدلة قرابة الوجه للمساعدة في تحقيقات الطب الشرعي بناءً على الشبكات العصبية العميقة

Verifying The Facial Kinship Evidence to Assist Forensic Investigation Based on Deep Neural Networks

Abstract

The human face contains a wealth of information that is influenced by genetics, therefore family members often share common facial characteristics due to their kinship. Kinship verification is a challenging problem in many important applications, such as forensic science, biometrics, and face recognition. Kinship verification provides a powerful tool in forensic investigations, contributing to the resolution of missing person cases, social media analysis, genealogy research, and historical study. Although a DNA test is the most accurate mean for kinship verification, it unfortunately, cannot be used in many scenarios such as situations that require real-time processing or applications where we have non-cooperative users, and it is also costly. The main aim of this thesis to verify if two people have a kinship by analyzing two face images together, extracting the relationship features between them, and then determining if they have Kin or not.
The proposed system presents a kinship verification system based on automatically learning discriminative features from facial images using a three-dimensional convolutional neural network with a new architecture. This system consists of two main stages: the preprocessing stage and the kinship verification stage, and each stage includes multiple steps that perform different functions. In the preprocessing stage, the input images are prepared to be suitable for deep neural network model by scaling and normalizing them. The kinship verification stage is implemented to provide the kinship decision in two steps: the feature extraction step to extract the salient facial features as well as tracking these features in two input images, and then the classification step to decide on those images: kin or not.
The extensive experiments revealed that the 3D CNN had promising results compared with many state-of-the-art approaches. The accuracy of the proposed system reached 83.75% in the KinFaceW-I dataset, 91% in the KinFaceW-II dataset, and 75.5% in the FIW dataset.