Research On Kinshipauthentication Algorithm Based On Feature Extraction And Metric Learning

The problem of kinship authentication based on images is to use computer vision method to determine whether the input pair of face images have a kinship.In real life,people often find that the facial features of parents and children have higher similarity than others,so parents' face images can be used to find missing children,build family genealogy,perform image search and annotation and so on.In this paper,three improved kinship authentication algorithms are proposed from the perspective of facial feature fusion and learning.First of all,based on how to use the local features of face images,a local feature fusion learning algorithm for neighborhood exclusion measurement was proposed.First,the key areas of the face were extracted,textures and skin color features were extracted for each key area,and then feature fusion was performed.Finally,metric learning was introduced.The transformation matrix was learned which can make the distance between the samples with kinship become smaller and the distance between the samples without kinship become larger.The extracted eigenvectors were mapped into the metric space through the transformation matrix,and the similarity was calculated by the cosine similar function.Secondly,based on how to use three objects for family relationship authentication,a local neighborhood exclusion measure learning algorithm with random combination of key features was proposed.It was that by comparing the Euclidean distances between the key facial features of the parents and the key facial features of the children,the key features with smaller Euclidean distances were selected as the approximate features of the key facial features of the children.Starting from the perspective of local features,each key feature was performed with metric learning respectively and then the cosine similar function was used to obtain the similarity of corresponding key features of each pair of samples.Thirdly,aiming at the problem that the traditional manual feature extractor is difficult to extract high-level abstract features,an End-to-End model of deep convolutional neuralnetwork was proposed.By iteratively optimizing the paired labeled training data into the network,the convolutional layer could extract the recessive features of the parent-child image pair,and the full-connected layer could map the extracted recessive features to the kinship-certified two-classification problem.The soft-max classification layer could directly determine whether the pair of samples has a kinship.