Research On Kinship Verification Based On Web Images

The ultimate goal of computer vision system is to acquire the adaptive ability,self learning abil-ity,the ability to balance the various solutions,the ability to use the new context and the ability of generalization and the ability to communicate with other systems(including people).Face,one of the important research objects in the field of computer vision,shows one prominent feature that it does not require aid from the visual subjects.With this advantage,face has drawn extensive attentions from the domains of pattern recognition and machine learning in recent several decades.After nearly thirty years of development,the face recognition system has begun to enter the business field from the labo-ratory environment.However,there are many different face recognition problems in the procession,for example,kinship verification based on Web images.Kinship verification based on Web images must deal with new problems and challenges due to the group images representation and verifier design.The group images representation problem comes from the complex appearance caused by great variations in imaging environment,expression,partial occlusion,pose and genetic characteristics.While the verifier design problem comes from the difficulty in group images representation,small sample problem and the large genetic difference.Because of the existence of these challenges,it is difficult to directly use the existing face verification algorithms to deal with the kinship verification problem.The focus of this paper is just towards robust kinship verification based on Web images.This paper focuses on three important topics involved in kinship verification based on Web im-ages,i.e.,kinship face representation learning,kinship verification model designing and promotion in practical application.To address the first topic,this paper proposes a spatially voted method for feature selection.For the second topic,a novel relative symmetric bilinear model is introduced to model the similarity between the child and the parents,by incorporating the prior knowledge that a child may resemble one particular parent more than the other.Finally,since the final aim is to promote kinship verification in practical application,a novel mixed kinship verification problem and the model is pro-posed.Specifically,the main contributions of this paper can be summarized as follows.(1)The problem of one-versus-two(tri-subject)kinship verification is proposed and a large-scale tri-subject kinship database characterized by over 1,000 child-parents families is released.Kinship ver-ification learning can be seen as a way to characterize the mutual infomation between multiple visual objects.However,existing kinship verification researches mainly focus on bi-subject relations,i.e.,father-son,father-daughter,mother-son and mother-daughter.It should be noted that more complex relations are involved in practical applications and the most basic relation among the human social re-lations is parents-son and parents-daughter relations.Understanding such kinship relations would have a fundamental impact on the behavior of an artificial intelligent agent working in the human world.Furthermore,it can promote the ultimate goal of computer vision,i.e.,transferring from characterizing single object to multiple entities.In addition,the one-versus-two kinship verification problem is eas-ier to implement than more complex kinship verification due to the scope involved in the problem is controllable and the problem itself is more easily defined.(2)A spatially voted method for feature selection is proposed.Under the framework of supervised kinship representation learning,discriminant and robust feature substraction method for kinship is dis-cussed.Aiming at the problem of the existing kinship representation learning,i.e.,considering each object separately,this paper takes local spatial information between kinship subjects into account and exploits those discriminative information for kinship verification.Particularly,all features extracted at each location in a given image are allowed to freely compete with each other and then the groups in which higher portion of individual features win are selected.Our proposed method works in a finer granularity than that of the group lasso,so higher performance can be obtained(3)A novel relative symmetric bilinear model is introduced to model the similarity between the child and the parents,by incorporating the prior knowledge of human sociology.The problem should be tackled in kinship verification is small sample problem.On the other hand,using auxiliary information is a powerful means to solve this problem.Activated by the researches of human sociology,a novel relative symmetric bilinear model is introduced by incorporating the prior knowledge hat a child may resemble one particular parent more than the other.Extensive experiments on TSKinFace and KinFaceW database show that the proposed method outperforms several previous state of the art methods,while could also be used to significantly boost the performance of one-vers-one kinship verification when the information about both parents are available.Furthermore,our proposed method can be thought of as a framework which can encompass any algorithm of bi-subject kinship verification for tri-subject kinship verification,while effectively incorporating useful prior knowledge.(4)A novel mixed bi-subject kinship verification problem and model designing method are pro-posed.Aiming at the problem involved in the existing methods,i.e.,extra gender annotation work may appear due to these methods are given based on the genders type,the promotion in practical applica-tion for kinship verification model is discussed and the mixed bi-subject kinship verification problem is introduced.Particularly,activated by the studies of human genetics,i.e.,such as eye or hair color or dimples have been proposed to be specifically linked to paternity confidence,we propose a novel multi-task learning method to address this problem with two transformation matrices-one is shared amongst all the tasks and the other is unique to each task.Both matrices are simultaneously learned in a joint framework,which enables our algorithm to utilize the common knowledge of the four tasks.Further,compared to single task learning,the advantage of multi-task learning is usually more visible when we only have few training samples per task.We expect that even if only few training samples are available for each task,the use of multi-task learning can compensate discriminative information by transferring from one task to the others.Lastly,we propose a multi-view multi-task learning method to perform multiple feature fusion to improve the mixed bi-subject kinship verification performance.