The Research Of Kinship Verification Based On Deep Learning

Kinship verification based on facial images mainly deals with verifying the kin relation between two subjects,namely verifying whether two persons have a biological kin relation or not by comparing their facial attributes.This new research topic has many potential applications such as family album organization,missing children searching.However,it has not been applied in our real life due to many extremely challenging problems.Three main difficulties in this task are the problem of extracting discriminative kin related facial feature descriptors,large inter-class similarity as well as intra-class difference and the scarcity of training data.Motivated by this,we approach this problem from the viewpoint of deep learning and have provided a novel large-scale kinship dataset to facilitate this verification task.Details about our algorithms and new kinship database are described as follows.Firstly,we have proposed a kinship verification method based on deep convolutional neural networks.Motivated by the powerful ability of data learning and feature representation in deep learning and considering the lack of a larger dataset available for kinship verification,we have developed a new verification algorithm utilizing deep learning for kinship verification on basis of the idea of transfer learning.Specifically,a deep convolutional neural network architecture is trained using facial images for face recognition,and then is employed as the feature extractor for kinship verification.Furtherer,the extracted face features are incorporated with metric learning in our algorithm,namely various distance metrics are learned using the deep features to reduce the similarity of facial image pairs without a kin relation and the difference of facial image pairs with a kin relation.Secondly,we have proposed a new kinship verification algorithm which is based on siamese networks.On basis of the siamese network and given the fact that kinship verification processes image pairs instead of a single image,our algorithm modified original siamese network and developed three basic convolutional neural architectures for kinship verification.The proposed algorithm takes facial image pairs as the input of the network by two channels in siamese network,and the whole kinship verification process becomes an end-to-end training system since the network is learned directly from the kinship data.In addition,the learned discriminative classification model improves the verification accuracy because contrasive loss utilized in our network reduces the distance between intra-classes and increses the ones between inter-classes to some extent.Lastly,we have introduced a new and large kinship database for kinship verification.Different from current kinship verification databases,our dataset has three advantages: including more data,cropping images in a positive kin pair from different original photographs and obtaining data under uncontrolled environments.These collecting facts,when taken into account,can significantly reduce the learning bias in kinship verification algorithm and are helpful for constructing a more discriminative classification model from the kinship dataset in verification algorithms.Furthermore,we have conducted massive experiments to eualuate the performance of the proposed approaches on these publicly available kinship databases.The comparison results have shown the superiority and effectiveness of our approach over state-of-the-art methods and human observers.