PARA

个性化图像美学评估 (PIAA) 由于其高度的主观性而具有挑战性。人们的审美趣味取决于多种因素，包括形象特征和主题特征。现有的PIAA数据库在注释多样性方面受到限制，尤其是在主题方面，已不能满足PIAA研究日益增长的需求。为了解决这一困境，我们对个性化图像美学进行了迄今为止最全面的主观研究，并引入了一个新的具有丰富属性的个性化图像美学数据库 (PARA)，该数据库由31,220图像组成，并由438主体进行注释。PARA具有丰富的注释，包括9个面向图像的客观属性和4个面向人类的主观属性。此外，还提供了脱敏的主题信息，例如人格特征，以支持PIAA和用户画像的研究。对注释数据进行了全面分析，统计研究表明，提出的主观属性可以反映审美偏好。我们还利用主题信息作为条件先验，提出了一个条件PIAA模型。实验结果表明，条件PIAA模型可以优于对照组，这也是首次尝试证明图像美学和主题特征如何相互作用以产生图像美学上复杂的个性化品味。我们相信数据库和相关分析将有助于进行下一代PIAA研究。

Personalized Image Aesthetics Assessment (PIAA) poses significant challenges due to its highly subjective nature. People's aesthetic preferences depend on multiple factors, including image-level features and subject-related features. Existing PIAA databases are limited in annotation diversity, particularly with regard to subject-related attributes, and can no longer meet the growing demands of PIAA research. To address this limitation, we conducted the most comprehensive subjective study on personalized image aesthetics to date and introduce a new personalized image aesthetics database with rich attributes (PARA), which consists of 31,220 images and was annotated by 438 individual subjects. PARA features rich annotations, including 9 image-oriented objective attributes and 4 human-oriented subjective attributes. Additionally, de-identified subject information (e.g., personality traits) is provided to support research on PIAA and user profiling. Comprehensive analysis was performed on the annotation data, and statistical studies demonstrate that the proposed subjective attributes can reflect aesthetic preferences. We further leverage subject information as a conditional prior to propose a conditional PIAA model. Experimental results show that the conditional PIAA model outperforms the control groups, which is also the first attempt to demonstrate how image aesthetics and subject-related features interact to generate complex personalized aesthetic tastes in image aesthetics. We believe that this database and the accompanying analysis will facilitate next-generation PIAA research.