Personalized Image Aesthetics Assessment via Meta-Learning With Bilevel Gradient Optimization

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Generic image aesthetics assessment (IAA) models fail to capture diverse user preferences.
• Personalized IAA (PIAA) is crucial due to significant variations in individual aesthetic judgments.
• Existing PIAA methods often rely on generic IAA models, neglecting aesthetic diversity.

Purpose of the Study:

• To develop a personalized image aesthetics assessment (PIAA) method that addresses the limitations of generic models.
• To learn shared prior knowledge from diverse users' aesthetic judgments.
• To enable rapid generalization to new users with limited data.

Main Methods:

• Proposed a novel PIAA method based on meta-learning with bilevel gradient optimization (BLG-PIAA).
• Employed a two-phase approach: meta-training and meta-testing.
• Trained an aesthetic meta-learner model using bilevel gradient updates on user-specific tasks (support and query sets).

Main Results:

• The BLG-PIAA method significantly outperforms state-of-the-art PIAA metrics.
• The learned prior model demonstrates rapid adaptability to unseen PIAA tasks.
• Experimental results validate the effectiveness of meta-learning for personalized aesthetics.

Conclusions:

• Meta-learning with bilevel gradient optimization offers a robust solution for personalized image aesthetics assessment.
• The proposed BLG-PIAA method effectively captures individual aesthetic diversity and generalizes well.
• This approach advances the field of personalized content understanding and recommendation systems.