A performance evaluation of convolution neural networks for kinship discernment: An application in digital forensics

Abstract

Kinship Verification from facial images is known to have attracted major attention since time immemorial. Identifying the underlying patterns that exist between images and analysing the relationship hidden between them have enabled the multitudes of applications to utilize kinship relationships. This work serves as a study on the amount of influence that hereditary features can exert on the families tied together by lineage in identifying the relationship prevailing between them and whether it holds true or not. The approach employed involves detecting the relationship existing between the provided facial images using Siamese Network, which comprises two identical convolutional neural networks that share common weight values. A difference vector is computed from this Siamese CNN, which is then fed into a network of fully connected linear layers. This extended layer will determine whether the two individuals in the input images are related to each other or not.