Building CNN-Based Models for Image Aesthetic Score Prediction Using an Ensemble.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Image Aesthetic Assessment (IAA) is crucial for understanding visual content.
• Existing methods for predicting image aesthetic scores (AS) have limitations.
• Convolutional Neural Networks (CNNs) show promise in image analysis tasks.

Purpose of the Study:

• To develop a framework for improving image aesthetic score (AS) prediction.
• To enhance IAA model performance through ensembling diverse CNN architectures.
• To analyze model attention regions for consistency with image subjects and photographic principles.

Main Methods:

• Constructed two types of IAA models utilizing different CNN architectures.
• Employed an ensemble approach to combine predictions from multiple models.
• Extracted model attention regions to investigate visual focus and interpretation.

Main Results:

• The ensemble method significantly improved AS prediction performance.
• Average F1 score of the ensemble increased by 5.4% over model type A and 33.1% over model type B.
• Models trained on the XiheAA dataset demonstrated learning of latent photographic principles.

Conclusions:

• The proposed ensemble framework effectively enhances image aesthetic score prediction.
• The study suggests that AI models can learn objective photographic principles from aesthetic datasets.
• Further research is needed to determine if true aesthetic sense can be learned.