Quantitative Analysis of Training Methods, Data Size, and User-Specific Effectiveness in DL-Based Personalized Aesthetic Evaluation

Yoshia Abe, Tatsuya Daikoku, Yasuo Kuniyoshi

Abstract: The perception of beauty is a highly subjective phenomenon that varies among individuals. In the past decade, advancements in deep learning technologies have led to numerous studies proposing AI models capable of predicting human aesthetic evaluations of images. Furthermore, several of these studies have focused on personalizing AI models for specific individuals. However, there have been challenges in collecting sufficient data required for deep learning. Recently, a large-scale image dataset for individual aesthetic evaluation has been proposed, enabling the quantitative analysis of the performance of deep learning-based personalized prediction models. The present study addresses this issue by focusing on training methods, data size, and effectiveness for individuals with different characteristics. The results demonstrate both the effectiveness and limitations of current deep learning-based training methods. Additionally, the analysis reveals the potential of AI as a promising entity that can deeply understand an individual’s aesthetic preferences. This study provides valuable insights and suggestions for developing superior personalized aesthetic evaluation models.

External IDs:dblp:conf/pricai/AbeDK24