Neighborhood min distance descriptor for kinship verification

The verification of parental and family relationships based on the facial appearance of subjects is a recent topic that attracted the attention of the computer vision research community. So many feature descriptors for facial images have been proposed, yet they are still unable to describe the similarity level of pairs of facial images in kinship datasets, because most of these images suffer from the illumination and expression variations. In this paper we propose a simple and effective descriptor for pairs of images, this descriptor is designed to be more robust to variations in expression, the idea of the descriptor is to compare each pixel in the first image of the pair with the neighboring pixels in the second image in terms of the euclidean distance on the RGB color space, the minimal distance in this neighborhood is then added to the feature vector of the pair. Experiments on the size of the neighborhood with various classifiers were conducted using the KinFaceW and Cornell-KinFace datasets, results demonstrated that this descriptor outperforms state-of-the-art approaches in five out of eight subsets of the KinFaceW and on the Cornell-KinFace dataset.