The Research On Key Problems Of Kinship Verification Based On Images

Image-based kinship verification determines the kinship relationship between parents and children by evaluating the similarity between the features of their facial images. As one of the main area in face recognition, it has attracted a lot of researchers and becomes increasingly popular with several potential applications. However, kinship verification is still a challenging problems. One of the main difficulties is the age gap between old parents and young children. Besides, the change of facial expressions and the accessories can also impact the local similarities evaluating. To handle the above difficulties, we have proposed two different kinship verification methods, namely, the transfer metric learning based kinship verification and the adaptive random pooling based kinship verification. Details are described as follows.Firstly, we have proposed a transfer learning based kinship verification method. The method introduces young parent images as an intermediate set to decrease the adverse impact of age gap between old parents and children. A metric space is learnt on parent set, young parent set and child set to minimize the differences between parent-child pairs with kinship relations and maximize the differences between those without. To further decrease the difference between the distributions of source and target domains, we have introduced an Adaptation Regularization support vector machine for kinship pairs classification.Besides, we have proposed an adaptive random pooling based kinship verification method. It believes that the random pooling scheme can avoid overfitting occurred in concatenating based pooling methods, and the information loss suffered in statistic pooling methods. In this method, face images are segmented into several patches. After their features are obtained respectively, the holistic representation is generated by randomly selecting some of the patches at different scales(i.e. different numbers of selected patches), and concatenating them to form a vector. In training stage, Adaboost is applied for each scale to adaptively adjust the sample weights and combine the advantage of weak classifiers to construct a strong one. In testing stage, the verification results on test dataset are obtained by the voting of results at different scales.To evaluate the effectiveness of the proposed methods, we carry out experiments on two publicly available dataset. Both of them are captured in natural environment. Comparison results with other popular methods indicate the effectiveness and robustness of our methods.