Human Relationship Mining And Recognition In Social Media

Social relationship is a peculiar phenomenon of human beings.With the advent of social media,the presentation of the relationship has become more intuitive.As a new branch of face analysis,image-based social relationship mining and recognition has attracted more and more researchers' attentions in recent years.However,in consideration of the effectiveness of dataset and its own difficulties,there is a gap between existing research achievement and practical application.As a result,image-based relationship recognition is still a challenging problem.In this thesis,we divide the relationship into kin and non-kin,and recognize them from buddy photos and family photos respectively.Accordingly,our own effective solutions on each task are proposed at the same time.The main content of this thesis is as follows.Firstly,literatures related to image-based friendship mining and kinship recognition are reviewed.Existing researches on each task are categorized according to the approach they used.Especially,representative works are highlighted,including their basic ideas and core algorithms.Secondly,for friendship mining task in social media using the recommendation system framework,a collaborative filtering motivated automatic photo tagging method is proposed.Compared to the item recommendations made by classic system for customers,the proposed algorithm predicts the probability of an identity's appearance in each photo.In accordance with the characteristic of social network photos,associated social context is also incorporated into a basic visual face recognition system.Experimental results demonstrate the effectiveness of our proposed model in identity recognition.Finally,on the basis of existing kinship verification and identification tasks,a pictorial structure based automatic family tree construction approach is proposed.Unlike the traditional works based on single image,the proposed album-oriented method is able to leverage the cooperation among multiple instances associated with a same identity.In this way,the reliability of kinship similarity measurement can be improved.Besides,a family tree structure is used to describe the kinship relations among each pair of members.Experimental results show that the proposed algorithm effectively improves the accuracy of kinship identification.