Kinship verification between two people by machine learning.

Abstract

The kinship verification field attracted much attention in the few past years due to its capacity to improve biometrics systems as a soft biometric for face verification (kinship traits) and an important role in many society applications (kinship verification). Among these applications include the creation of family trees, family album organization, image annotation, finding missing children and forensics. Although, the DNA test is the most trustworthy way for kinship verification, it cannot be used in many situations. Automatic kinship verification from facial images can exemplary be done in video surveillance scenes. In this thesis, facial kinship verification over facial images is studied. At this end, we start with the previously proposed approaches like features learning-based kinship verification methods, metric learning-based kinship verification methods, and convolutional deep learning-based kinship verification methods. Also, the general facial kinship verification system is presented, challenges and measures of characteristics are mentioned. Furthermore, the various evaluation terms are illustrated. Concluding with the proposed approaches and the obtained results on various databases. The proposed frameworks comprise of three main phases as follows: 1) features extractions; 2) subspace transformations analysis; 3) kinship verification decision. The aim of feature extraction is to extract discriminative representations of facial images. This phase is important since the kinship traits are very sensitive to the unconstrained environments (i.e. facial images captured under uncontrolled environments without any restrictions in terms of pose, lighting, background, expression, and partial occlusion). Also, it can affect the final decision performance of the framework. Subspace transformations analysis phase extract and select the more attractive and discriminative facial traits. Therefore, the features are extracted by a projection of the original data (features) of the previous phase to get better discrimination and make more precise decisions. In the last phase, cosine similarity is used as the best metric compatible with discriminant analysis methods (subspace transformations analysis methods) and kinship verification. The final metric between two facial images is compared to a threshold to decide if the pair facial images come from the same family or not. Finally, our results show great improvement for facial kinship verification on the largest and smallest databases. Also, a robust and good performance was achieved by the proposed systems and comparing favorably with the state of the art approaches. The proposed frameworks are also convenient for real-time applications.