Kinship Mining And Recognition In Social Media

Kinship is one of the major social relationships in social media.Recently,image-based kinship mining and recognition is becoming an important area in social media analytics,and has gained much interest from researchers.However,there are many drawbacks in existing related works,and researchers in this field still face huge challenges.In this paper,two specific tasks of kinship mining and recognition are explored,i.e.,tri-subject kinship verification between a child and a couple,and family photo recognition(i.e.,classifying family and non-family photos).In each task,we review related works and find their problems,and then propose our method to deal with the problems.The main works of this thesis are as follows.Firstly,related works in kinship recognition and family photo recognition are reviewed respectively.In each task,related works are categorized based on their core methods,and their ideas and core mathematical models are introduced.Secondly,a genetics-motivated tri-subject kinship verification algorithm is proposed.Pre-vious verification models do not accord with the inheritance process of facial features.In this thesis,based on the genetic process from two parents to a child,the kinship dissimilarity is computed between the linear feature combination of two parents and a child's feature.In addi-tion,a multi-scale,high-dimensional representation is applied to extract genetic facial features.Experimental results demonstrate that the proposed method achieves good performance in both accuracy and speed.Finally,a family photo recognition algorithm based on multiple instance learning and local features is proposed.Previous methods fail to utilize the local discriminative information in group photos.Accordingly,in this paper,three local representations are proposed to extract the kinship,geometry and semantic information which lies in local parts of a group photo.Besides,multiple instance learning is used to describe the multiple local features and do the classification.Experimental results verify that the proposed method achieves better recognition performance than previous methods.