Social-sensed Image Personal Aesthetics User Modeling And Identification

The computational image aesthetics is to endowing computer with the ability to assess the aesthetics value of images as human beings does.Nowadays,most of the scholars’ work on image aesthetics is based on the common sense of human beauty which distinguishes high-aesthetic images from low-aesthetic ones.Although the perception of beauty for human beings is in common,personalities are still strongly affecting the final result.Thus,our works on user’s image personal aesthetics will be very important for image aesthetics,especially in image recommendation.With the vigorous development of social images,considerable semantic annotations and behavioral data can be preserved,which provide a strong support for the user’s personal image analysis.So far,only the visual information of the image was considered in the study of personal aesthetics by researchers,without making use of social information of the image.In this paper,we mainly study on the expression and modeling of user’s personal aesthetics with social images.By extracting the visual features of the user’s images and introducing the semantic information of the social image,we employed a 2 dimensional generative model to capture the user’s personal tastes,after which we presented a soft biometric approach based on social-sensed image personal aesthetics in identification tasks.We focus on the user’s image personal aesthetics with social images,and our works are as follows:1)An image dataset including 200 users and 40,000 images named SCUTUserData was created for personal aesthetics research based on users’ preferred images from the Flickr community.2)We realized the image visual feature extraction in SCUTUserData image dataset,including the aesthetic features and the content-based features.3)We had designed the tag features for images by using the tags added by the uploader from image social networks,and finally we combined the tag features with the visual features of the image as the input of the generative model.4)We had designed the weight of users’ images to weight the popular images and not popular images by considering the interactive information of the image in social networks,the views and faves of an image.In particular,the not popular images were weighted higher to capture the personal tastes,while the popular images were weighted lower,since these images represent the common sense of beauty for human beings.In this paper,a generative model,Counting Grid,was employing to capture users’ personal aesthetics and then SVM was trained for every user to conduct user identification based on social-sensed image personal aesthetics.We used 100 preferred images as train and test samples for each user from SCUTUserData image dataset giving 84% of probability of guessing the correct user,which is the state-of-the-art in personal aesthetics for user identification and significantly impored the efficiency of time.