Personalized Image Aesthetics Assessment Based on Theme and Personality

Personalized Image Aesthetic Assessment (PIAA), which studies individual users’ aesthetic preferences for images. Although having achieved good performance on a few image aesthetic assessment datasets, the existing PIAA methods still have some shortcomings. They directly learn aesthetic features from images without considering the influence of theme changes. And due to the lack of user annotation aesthetic assessment data, it is hard to adopt those models to new users. As evaluation criteria vary for images of different themes and aesthetic preferences differ among users with different personalities, we propose a PIAA method which simultaneously integrates the information of aesthetic attributes, theme information and user personalities. To capture the correlation between aesthetic attributes and user personalities, we introduce a multi-task network to evaluate personalized aesthetic scores by learning aesthetic features and personality traits, and we also add a theme feature extraction network to improve the robustness of aesthetic score prediction. Experimental results demonstrate that our approach outperforms other methods in learning the personalized image aesthetics of individual users.