A Research On Personalized Image Aesthetics Assessment And Representation

Image aesthetics assessment,or computational aesthetics,is a technology to summary user's image aesthetics preference or give an asseessment for a specific image by computer following human's visual and psychological cognition.Recently,with the breakthrough of deep learning,great progress for a generic image aesthetics assessment model has been made.However,image aesthetic preference is very subjective,different users may have different perceptions.Thus,a generic image aesthetics model may not suitable for every user.It's urgent for the research of personalized image aesthetics.Personalized image aesthetics is a research based on computable image aesthetics,which considers each user individually and learns the user's image aesthetic preferences.This research is mainly faced with the following challenges.First,the lack of public personalized image aesthetics dataset.Not only of the difficultly in collecting personalized information,but also the small amount of data from one individual user.Secondly,such user-specific image aesthetics is abstract without any clear rules,which is same as image aesthetics assessment.Such challenges had slowed down the progress of related research and made it relative lack of references.In this paper,two specific tasks in the research of personalized image aesthetics had been considered.User aesthetic assessment,which studies how to build an image aesthetic assessment model that matches their preference for individual user;User aesthetic representation,which studies a how to represent user in a better way by their favored images and used it for user identification and recommendation.Although meaningful pioneering works have been done,there are still some problems in the existing research work:1.For user aesthetic assessment,existing methods use aesthetic attributes or users' online interaction and build a 2-stage residual model.However,they did not solve the problem of too little user data,and the information extracted from user data is limited.2.For user aesthetic representation,existing methods use a generative model the build it.They tend to build a sparse structure,reduces the information density of user representation and makes it inefficient.Considering the aforementioned problems,this paper takes the personalized image aesthetics as the research object,mainly focus on user aesthetics representation and user aesthetics assessment.This paper has made the following innovative achievements:1.For user aesthetic assessment,we firstly introduce meta learning into user aesthetic assessment to deal with the little amount of data from individual user.We also proposed a novel meta learning strategy and meta regularization term for a better generalization performance.Our method could give a more precise personalized image aesthetic assessment,which is more consistent with the user's real score.2.For user aesthetic representation,we proposed a gird space aggregation strategy in a generative model,turned the original sparse user representation into a dense way,with a bigger information density and more consistent with numeric calculation.We also made an improvement for this aggregation aims at data imbalance case in real situation.