Monday 31 March 2025
The quest for a more personalized image aesthetic assessment has been ongoing in the field of computer vision. Researchers have been working towards developing models that can accurately predict how individuals perceive and appreciate images, taking into account their unique characteristics and preferences. A recent study published in the journal Brain and Cognition sheds light on this topic, exploring the role of individual differences in current approaches to computational image aesthetics.
The authors propose a new model called NIMA-trait, which utilizes a ResNet-50 architecture as an image encoder. The model is trained on two datasets: PARA, which consists of photographs, and LAPIS, which features artworks. The researchers found that the model’s performance varies significantly depending on the demographic characteristics of the users, such as age, education level, and artistic experience.
One of the key findings of the study is that individual subjectivity plays a crucial role in determining how images are perceived and appreciated. The authors demonstrate that averaging individual scores does not eliminate individual differences, which can lead to suboptimal performance when evaluating images on unseen users. This highlights the importance of incorporating personal traits and characteristics into image aesthetic assessment models.
The researchers also investigate the transfer learning between generic image aesthetic assessment (GIAA) models and personalized image aesthetic assessment (PIAA) models. They find that GIAA models tend to perform better than PIAA models, but only when evaluating images on users from the same demographic group as the training data. This suggests that GIAA models are less robust to unseen users.
The study’s findings have significant implications for the development of image aesthetic assessment models. The authors propose a unified model that encodes individual characteristics in a distributional format, allowing it to be used for both individual and group assessments. This approach can help improve the model’s performance on unseen users by taking into account their unique traits and preferences.
The researchers also demonstrate the effectiveness of their proposed model on two datasets, PARA and LAPIS. They show that the model outperforms existing state-of-the-art models in terms of similarity to human judgments. The study’s results highlight the importance of considering individual differences when developing image aesthetic assessment models and underscore the need for more personalized approaches.
The authors’ work provides a valuable contribution to the field of computer vision, shedding light on the complex interplay between individual characteristics and image perception. As the demand for personalized content recommendation systems continues to grow, this research has significant implications for the development of more accurate and effective image aesthetic assessment models.
Cite this article: “Personalizing Image Aesthetic Assessment: Uncovering Individual Differences in Computational Models”, The Science Archive, 2025.
Computer Vision, Image Aesthetics, Personalized Assessment, Individual Differences, Subjectivity, Transfer Learning, Generic Image Aesthetic Assessment, Personalized Image Aesthetic Assessment, Demographic Characteristics, User Traits.
