Feature study and analysis for unseen family classification

Abstract

Due to the genetic proximities, siblings are often observed to bear close facial resemblances to each other as well as their parents. In this paper, we attempt to develop such human capability in computers. In order to achieve this goal, Haar, Gabor, SIFT and SURF features of family and nonfamily datasets are extracted and used for AdaBoost to train the classifier. The primary difference between our study and other relevant applications like face recognition, album auto tagging and annotation is that the query person we intend to classify may not even exist in the training data. We have conducted testing for various scenarios where different members of the family are absent from training but present in testing, and have obtained interesting results with practical implications for the development of automated family member classification. As family data sets used in this paper has good quality colour samples, we use FERET dataset as non-family samples to have fair comparison. Results obtained show that we can achieve up to 87% accuracy depending on the absent family member.