Research On Key Point Of Kinship Verification Using Neural Network

As an interesting and challenging new branch in image processing and computer vision,kinship verification has a wide range of potential applications in many fields,including searching for missing children,constructing family photo albums,social media analysis,and automatic labeling of pictures.In recent years,kinship recognition has received extensive attention from the academic community and has made great progress.The kinship verification algorithm generally includes two modules: feature extraction,and the classification of relatives.Usually these two phases are separate,and the extracted features may not be discriminative to the classification.At the same time,the individual features are not sufficiently comprehensive for the representation of the relative images.To solve the above problems,in this paper we first propose an end-to-end learning model using neural networks,which combines the feature extraction and classification stages to form a taskdriven network.At the same time,a new loss function is added to constrain the features.Then,in order to further establish a more discriminating relative feature representation,we propose to establish a neural network model based on multi-feature fusion,that is,to optimize multiple features.Details are as follows:(1)This paper proposes a kinship verification algorithm using an angle-loss neural network.This algorithm is based on the convolutional neural network to extract depth features of target image pairs.Unlike the general method of extracting manual design features,this algorithm learns more discriminative features of classification through the end-to-end convolutional neural network based on task learning.The target's feature information is more abundant.In addition,we added an angle loss function to the algorithm to make it more robust to the problem of similar facial features among non-relatives at the time of classification.By using the angle loss function to map the features used for classification from the Euclidean space to the hypersphere,this is more in line with the distribution of the face on the nonlinear manifold domain,and the distance between classes is made smaller by the restriction of parameters.The distance between classes is greater so that learning more discriminative classification features is more conducive to the classification of kinship,the experimental results on the kinship verification dataset verify the effectiveness of the algorithm.(2)This paper then proposes a kinship verification algorithm based on multi-feature fusion neural network.The algorithm performs a deep-level non-linear fusion of facial feature representations extracted by commonly used feature extractors to obtain more information and more robustness relative identification features.The feature extractor may be a traditional manually designed low-level feature,or it may be a deep high-level feature acquired by a deep neural network.Features are more effective for solving certain problems in relative identification,such as lighting and posture transformation.Fully connected layers and nonlinear activation layers through neural networks can effectively fuse the similarity and complementarity of multiple features,so that the neural network can learn the fusion of various features,the common enrichment of semantic features at each level.The experimental results on the kinship verification dataset verify the effectiveness of the algorithm.