Yang Chang¹, Denise Hsien Wu¹; ¹National Central University, Taiwan

What makes certain images more pleasurable than others? While much research has focused on low- and mid-level visual features (e.g., symmetry, complexity), the impacts of high-level features (e.g., semantic properties) and information processing characteristics (e.g., perceptual organization) on sensory pleasure remains largely unexplored. Moreover, previous research often employed limited and oversimplified stimulus sets, failing to account for the multidimensional nature of human visual experience. These caveats have led to inconsistent findings across different studies, hindering a generalized view of the factors contributing to sensory value construction. In the present study, we selected nearly 7,000 complex images, including everyday photographs and paintings, from existing datasets. These images are both naturalistic and multidimensional, and have been rated based on various attributes (i.e., beauty, enjoyment, liking, and valence) relevant to pleasure experience. Leveraging on computational analysis and modeling with vision and language-based neural networks for feature extraction, we identified over 200 image features, categorized into three groups: visual statistics, information processing, and semantic properties. Based on these features, we built regression models to predict image pleasure ratings and examined the contribution from each feature category to the prediction. The results demonstrated that for photographs, semantic properties had the strongest influence on pleasure ratings, followed by information processing, while the order was reversed for paintings, with information processing exerting the strongest influence. Visual statistics had the smallest but still significant impacts on both image types. As for the ratings of valence, semantic properties showed a stronger effect than the other feature categories. Further analysis of semantic properties revealed that object identity (nouns) contributes more to sensory pleasure than object relation (verbs) does. In summary, these findings provide a comprehensive understanding of the features driving human preference for high-dimensional images, emphasizing the crucial role of conceptual representation and perceptual organization in shaping sensory valuation.

Acknowledgements: This work was funded by National Science and Technology Council, Taiwan. grant ID: NSTC111-2410-H-008-060-MY3