Study On The Personality Analysis And Personalized Image Aesthetics Assessment Of Users

With the continuous improvement of material living standards,people’s spiritual and cultural demands are also increasing.At present,the main contradiction in our society has been transformed into the contradiction between the people’s growing need for a better life and the unbalanced and inadequate development.In recent years,with the rapid development of mobile internet and social networks,multimedia contents(e.g.,images)have exploded,which meets people’s needs for a better spiritual life to a certain extent,such as beautiful food,art photos,film,and television.Therefore,it is more and more urgent to study the image content that meets people’s visual aesthetic demands,which is of great significance to the construction of social spiritual civilization at this stage.Image aesthetic assessment is to study how to use intelligent computable technology to imitate people’s aesthetic perception of images and automatically predict the results of aesthetic evaluation.In this way,we can screen out the image content that meets people’s aesthetic demands.However,since people’s visual aesthetic is highly subjective,and it is often affected by personal psychological state and emotional factors,the evaluation of image aesthetics is a very challenging task.At present,most of the research on image aesthetic assessment mainly focuses on the generic image aesthetic assessment,which is to predict the results of most people’s aesthetic evaluation of images.However,these methods do not take into account the subjective characteristics of image aesthetics,and cannot accurately evaluate the personalized image aesthetics results of specific users.In recent years,although some personalized image aesthetic assessment methods are proposed for users,these methods mainly use the objective attributes of the image to model the user’s personalized aesthetic perception,ignoring the important role of the user’s subjective characteristics(such as personality traits and emotions)in their visual aesthetic perception.Hence,this paper studies the abovementioned problems of users in the aesthetic perception of images.First,the users’ aesthetic preference on images in social networks is used to analyze and study the user’s personality traits,and the approaches for predicting user personality traits based on liked images are proposed.Then,we take the user’s personality traits as an important subjective feature,assisting the image aesthetics assessment method,and obtain a personalized image aesthetics assessment model that conforms to the user’s subjective perception.Finally,in view of the small sample properties existing in the evaluation of personalized image aesthetics,the meta-learning method is used to build the aesthetic prior meta-knowledge model for quickly adapting to the target user’s personalized image aesthetics assessment task.Therefore,exploring the personality analysis and personalized image aesthetics assessment methods of users in social networks is an urgent problem in the development of social spiritual civilization,and it is also an important research topic in the fields of psychology,computer vision,and artificial intelligence.This dissertation focuses on the analysis of users’ subjective aesthetic preference of images in social networks,from the perspectives of users and images.Firstly,the users’ aesthetic preference on images is used to predict their personality traits,and then the personalized image aesthetic assessment methods for specific users are studied.The main research content is divided into the following four aspects:(1)The image scene is a highly abstract semantic feature.Users have different preferences for the image scene,which reflects their different personality traits.Inspired by the relationship between the user’s personality traits and their preferred image scenes,this paper proposes a method for predicting user personality traits based on the probability of scene perception.Firstly,the scene recognition method based on a convolution neural network is used to recognize the scene of the user’s preferred images;then,the scene probability distribution of the user’s preferred image is further generated.Finally,the linear regression model is built by using scene probability distribution to predict the user’s personality traits.Experimental results on the Pshyco Flickr dataset show that the proposed method can effectively predict the user’s personality tarts through the scene information of their liked images.(2)The user’s visual attention to the image usually has the local attention characteristic,so this characteristic should be considered in the personality trait prediction method based on users’ preferred images.In this paper,a weak supervised dual convolution network is proposed to predict the user’s personality traits,which is composed of a classification network and a regression network.Among them,the classification network can capture the attention region features of users with different personality traits when only using the image-level user personality category label for supervision training.In the regression network,the user’s personality traits are further predicted by combining the global features and local attention features of the preferred images.Therefore,this method can not only predict the user’s personality traits from their preferred images but also obtain the attention region features of users with different personality traits.The experimental results on the Pshyco Flickr dataset show that the proposed method is superior to the state-of-the-art personality prediction methods based on users’ preferred images.(3)The traditional image aesthetics assessment method mainly evaluates the average image aesthetics of most people.However,different users have different aesthetic preferences for images,which is mainly determined by their different subjective visual preferences.As an important subjective characteristic,personality traits are a key factor in modeling users’ subjective preferences.This paper proposes a personalized image aesthetic assessment method based on personality-assisted multitask learning.The proposed method framework comprises two stages.In the first stage,a multi-task learning network with shared weights is proposed to simultaneously predict the aesthetic distribution of an image and the personality traits of users who prefer the image.In order to obtain the common representation of the generic aesthetics of the image and the user’s personality traits,this paper constructs a Siamese network and uses the aesthetic data and personality data to jointly train the multi-task learning network module.In the second stage,based on the user’s personality traits and generic aesthetic scores predicted in the multi-task module,the inter-task fusion learning module is further introduced to finally generate a personalized image aesthetic assessment model for specific users.The performance of this method is evaluated on two public image aesthetics databases.The experimental results show that the proposed method is superior to the state-of-the-art methods in both generic and personalized image aesthetics assessment tasks.(4)Since it is difficult to obtain a large number of samples of aesthetic annotations on images by specific users in real life,the user’s personalized image aesthetic assessment is a typical small sample-learning problem.In response to this problem,the existing personalized image aesthetics assessment model is generally obtained by finetuning the generic image aesthetics assessment model as prior knowledge.However,this prior knowledge based on generic average aesthetics cannot reflect the diversity of image aesthetics of different users.In order to learn the prior knowledge shared by different users in the aesthetic evaluation of images,this paper proposes a personalized image aesthetic assessment method based on gradient optimization-based metalearning.This method can directly train a large number of users’ personalized image aesthetics assessment tasks to build a prior knowledge model and then fine-tune a small number of training samples of target users for quickly adapting to the user’s personalized image aesthetics assessment tasks.The experimental results show that the performance of the proposed method outperforms the state-of-the-art personalized image aesthetics assessment methods.