LR Coarse-to-Fine Image Aesthetics Assessment With Dynamic Attribute Selection (Huang et al., 2024)

Coarse-to-Fine Image Aesthetics Assessment With Dynamic Attribute Selection (Huang et al., 2024)

1. Overview

This paper addresses the problem of Image Aesthetics Assessment (IAA), which aims to predict the aesthetic quality of images. Unlike traditional approaches that directly map images to aesthetic scores, the authors propose a human-like reasoning framework that performs coarse-to-fine aesthetic evaluation. The proposed model, CADAS, integrates attribute learning, dynamic attribute selection, and feature fusion to improve both prediction accuracy and interpretability.

---

2. Problem Definition

Existing IAA methods typically suffer from three limitations:
(1) they directly regress aesthetic scores without mimicking human perception,
(2) they fail to consider the varying importance of aesthetic attributes, and
(3) they lack explainability, providing only scores without reasoning.
This paper aims to address these issues by modeling the staged human aesthetic perception process and identifying dominant attributes.

---

3. Key Idea

The core idea of this work is to simulate human aesthetic judgment through a coarse-to-fine pipeline.
First, the model performs a coarse binary classification (aesthetically pleasing or not). Then, it dynamically selects the most influential aesthetic attributes based on classification confidence. Finally, it combines these attributes with visual features to predict fine-grained aesthetic scores and distributions.

---

4. Method

The proposed CADAS framework consists of three main components:

• AttributeNet:
  A hierarchical network that predicts aesthetic attributes (e.g., color, lighting, composition) in a coarse-to-fine manner, inspired by human perception.
• AestheticNet:
  Performs binary classification to estimate whether an image is aesthetically pleasing or not. This step provides a confidence score used for attribute selection.
• Dynamic Attribute Selection:
  Selects a subset of dominant attributes based on classification confidence. The number of positive and negative attributes is determined by predefined thresholds.
• FusionNet:
  A self-attention-based module that fuses selected attributes with visual features to predict the final aesthetic score distribution.

Pipeline summary:
image → attributes → coarse classification → attribute selection → fusion → score/distribution

---

5. Contributions

• Proposes a coarse-to-fine framework for IAA that aligns with human perception.
• Introduces a dynamic attribute selection mechanism to identify dominant attributes.
• Improves model explainability by explicitly outputting influential aesthetic factors.
• Achieves state-of-the-art performance on multiple IAA benchmarks.

---

6. Strengths

• Effectively models human aesthetic reasoning process.
• Provides interpretable outputs via dominant attributes.
• Demonstrates strong performance across classification, regression, and distribution tasks.
• Modular design allows flexible integration with different backbones.

---

7. Limitations

• The attribute selection mechanism relies on heuristic threshold rules rather than being fully learnable.
• Model performance depends on the quality of attribute prediction.
• Additional complexity due to multi-stage architecture (AttributeNet + AestheticNet + FusionNet).
• May not generalize well if attribute annotations are unavailable or noisy.

---

8. Insights and Future Directions

• The dynamic selection process could be replaced with a learnable attention-based selection mechanism.
• Incorporating multimodal information (e.g., text descriptions or user preferences) may further improve aesthetic understanding.
• The explainability framework could be extended to other vision tasks such as image quality assessment or recommendation systems.

---

9. One-line Summary

A coarse-to-fine, attribute-aware framework that dynamically selects dominant aesthetic factors to achieve accurate and explainable image aesthetics assessment.