Research On Kinship Recognition Method Based On Face Images

Kinship recognition technology based on face images is a kind of computer technology that can make full use of the facial biometric information and excavate the kinship relationship between the characters in the static images or videos through machine learning methods.As a new branch of face image analysis,it has become a research hotspot which could be widely used in social media analysis and information mining,searching for missing people,image annotation and retrieval.Based on the existing face recognition algorithms,this thesis focuses on the method of kinship verification based on face images.Firstly,in order to solve the problems that the traditional feature descriptors are not enough to describe the kinship relationship between the characters in the picture and a certain amount of redundant information in the feature vector is not conducive to measuring the similarity between relatives,a new kinship verification method based on deep transfer learning and feature nonlinear mapping is proposed.In this method,the high-level feature informations are firstly obtained from the kinship recognition datasets by transferring the deep learning model,and then mapped in a non-linear way by combining the siamese multi-layer perceptron with the method based on triangular similarity metric learning.Finally,the cosine distance between each sample pair is adopted to SVM(Support Vector Machine)for classification.The experimental results indicate that the proposed method improves the recognition rate compared with the traditional feature methods,deep learning methods and metric learning methods.Secondly,in order to solve the problem that the feature non-linear dimensionality reduction method only considers the kinship relationship between specific sample pairs and ignores the kinship relationship between specific example pairs with other samples,and in order to highlight the local similarity in the facial region between relatives and make full use of the complementaritary information between multiple types of features,a novel kinship verification method based on triangular similarity deep metric learning and multi feature fusion is proposed.In this method,multiple types of features are firstly extracted from the whole face image and the face image patches cut by the key regions,and then mapped in a nonlinear way by the method based on triangular similarity deep metric learning.Finally,a variety of cosine distances of each sample pair are adopted to the BP(Back Propagation)neural network for feature fusion and classification.The experimental results indicate that the proposed method improves the recognition rate compared with the method based on triangular similarity metric learning and the single feature methods.Finally,in order to meet the needs of kinship recognition,the age classification model and gender classification model are trained,and a kinship recognition prototype system based on face images is developed which integrates extracting feature methods,feature nonlinear mapping,kinship verification,age classification,gender classification and kinship recognition.The system can excavate the kinship relationship between the characters in the static face image pairs.