DiscoStyle: Multi-level Logistic Ranking for Personalized Image Style Preference Inference

Learning based on facial features for detection and recognition of people′s identities, emotions and image aesthetics has been widely explored in computer vision and biometrics. However, automatic discovery of users′ preferences to certain of faces (i.e., style), to the best of our knowledge, has never been studied, due to the subjective, implicative, and uncertain characteristic of psychological preference. Therefore, in this paper, we contribute to an answer to whether users′ psychological preference can be modeled and computed after observing several faces. To this end, we first propose an efficient approach for discovering the personality preference related facial features from only a very few anchors selected by each user, and make accurate predictions and recommendations for users. Specifically, we propose to discover the style of faces (DiscoStyle) for human′s psychological preference inference towards personalized face recommendation system/application. There are four merits of our DiscoStyle: 1) Transfer learning is exploited from identity related facial feature representation to personality preference related facial feature. 2) Appearance and geometric landmark feature are exploited for preference related feature augmentation. 3) A multi-level logistic ranking model with on-line negative sample selection is proposed for on-line modeling and score prediction, which reflects the users′ preference degree to gallery faces. 4) A large dataset with different facial styles for human′s psychological preference inference is developed for the first time. Experiments show that our proposed DiscoStyle can well achieve users′ preference reasoning and recommendation of preferred facial styles in different genders and races.