Study On Relation Network For Zero-shot Learning And Kinship Verification

As an important method of deep metric learning,Relation Network(RN)has been widely used in image classification,relationship analysis and other tasks.In essence,the purpose of the Relation Network is to train a deep neural network to automatically find the appropriate feature space and distance metric function,so as to realize the similarity judgment between samples.The research of Relation Network model is mainly divided into two aspects: network structure and model parameters.By analyzing the characteristics of data,the former can be solved well.While the latter can be realized only by designing optimization algorithm specifically.Usually,when an appropriate Relation Network model is designed according to the characteristics of the problem,the success of model parameter optimization depends on whether the learning objective is consistent with the task,that is,an appropriate loss function is designed to measure the difference between the network output and the real label.In this paper,the Relation network is used to solve the Zero-shot Learning task and the kinship verification task.And the research on the structure and parameter updating of the Relation Model can be well studied according to the task characteristics.In practical application,Zero-shot Learning task can identify categories that have never been seen before,and such idea can effectively solve the problem of data miss.Kinship Verification plays an important role in the identification of missing persons and the analysis of social media.Both of them have high potential research value.For the Zero-shot learning task,the relationship between visual image and category text description can be established by the Relation Network,and then the category of samples can be decided according to the distance of the relationship.In this paper,Zero-shot Learning is considered as a cross-module metric learning problem and a depth measurement model named class prototype discriminant network is presented.At the same time,in order to obtain a discriminative potential space,two novel losses are presented named Prototype-Sample Metric Loss and Class-Prototype Scatter Loss,so as to enhance the similarity between the prototype and the homogenous samples,reduce the similarity between the prototype and the non-homogenous samples and maintain the difference between the prototypes,and finally improve the performance of the Relation Network in the Zero-shot Learning.For the kinship verification task,Relation Network can be used to analyze different face features and decide whether the given face image pair has the kin relationship or not.We present the adaptively weighted 6)-tuple metric network in this paper.It can construct discriminative visual features by combining the features learned from different convolutional layers.At the same time,we also present a new cross-pair metric learning loss for networks: adaptive weighted 6)-tuple loss.Such loss provides an effective and smooth way to exploit the high order cross pair information for better maximizing the margin between positive and negative pairs.Moreover,its adaptive weighting scheme also provides an automatic way for selecting hard negative pairs to further enhance the metric learning and finally obtains an effective kinship verification model.