Abstract
Color temperature, as a crucial attribute influencing image color, plays a critical role in Image Aesthetics Assessment (IAA). Yet, within the existing IAA field, little light has been shed on assessing the aesthetic quality of image color temperature. To bridge this gap, we introduce a new task: Image Color Temperature Aesthetics Assessment (ICTAA). However, this task poses the following challenges: 1) Perceptual Sensitivity: humans exhibit high sensitivity to subtle shifts in color temperature, necessitating a model to enable fine-grained discrimination; 2) Spectral Continuity: The theoretical modeling of color temperature aesthetics requires continuous labels; however, the just-noticeable-difference property of human perception makes continuous labeling infeasible, necessitating a well-designed labeling strategy. To address the aforementioned challenges, we make the following efforts. First, we propose a multi-modal contrastive learning framework, ICTA2Net, that models color temperature differences between image pairs while strictly controlling other visual attributes. Second, leveraging color temperature transitivity, we design a weakly supervised strategy that discretely samples images based on anchor images and human perception to build contrastive relations across color temperatures, enabling learning from discrete labels. Thirdly, we construct a color temperature aesthetics dataset, ICTAA240K, and a benchmark for validation. Additionally, we propose a new metric, Information Entropy-weighted Accuracy (IEA), which weights accuracy by the degree of annotation disagreement to reflect model performance across varying sample difficulties, complementing existing evaluation metrics. Experiments show our method outperforms existing state-of-the-art IAA methods on ICTAA240K, thereby setting an effective roadmap for ICTAA.
Introduction
While Image Aesthetics Assessment (IAA) has traditionally focused on composition and lighting, the impact of color temperature—the perceived "warmth" or "coolness" of an image—has been largely overlooked. Color temperature is a fundamental driver of mood and aesthetic quality in photography. To address this, we introduce the Image Color Temperature Aesthetics Assessment (ICTAA) task.
The task presents two significant challenges. First, Perceptual Sensitivity: humans are highly sensitive to minor color temperature shifts, requiring models with fine-grained discriminatory power. Second, Spectral Continuity: while color temperature is continuous (Kelvin), human preference is subjective and non-linear, making continuous labeling impractical.
Methodology: ICTA2Net
To tackle these challenges, we developed ICTA2Net, a multi-modal contrastive learning framework designed to model color temperature differences between image pairs.
The architecture consists of several key components:
- Color Temperature Encoder (CTE): Extracts features specific to color shifts while filtering out irrelevant noise.
- Contextual Awareness Module (CAM): Employs a Text Denoise Model to refine image descriptions, allowing the model to understand the scene's context (e.g., a "sunset" vs. a "glacier").
- Cross-Modal Fusion Module (CFM): Aligns visual features with refined text to capture the relationship between theme and color.
- Pairwise Ranking Predictor (PRP): Predicts which of two image versions is more aesthetically pleasing based on learned features.
Weakly Supervised Learning
To overcome labeling difficulties, we utilize a weakly supervised strategy. By using professionally edited "anchor images" and generating shifts in both directions, we leverage the principle of transitivity to automatically create millions of training pairs, enabling the model to learn from discrete labels.
Dataset and Evaluation
We introduce ICTAA240K, a large-scale dataset of over 240,000 images derived from the MIT-Adobe FiveK and PPR10K datasets. Images were rendered at various temperatures (2000K to 10000K) using the Adobe Camera Raw SDK.
To better evaluate model performance on subjective tasks, we propose Information Entropy-weighted Accuracy (IEA). Unlike standard accuracy, IEA weights samples based on annotator agreement. This provides a more nuanced view of model reliability, especially on "hard" samples where human experts might disagree.
Results and Discussion
ICTA2Net was benchmarked against 10 state-of-the-art IAA models across four specialized subsets (General and Portrait scenes). Our model consistently achieved the highest scores. Notably, on portrait scenes (ICTAA-HP), ICTA2Net improved accuracy by 9.3% over the next best model.
Ablation studies confirmed that the inclusion of textual context is vital; removing text descriptions significantly degraded performance, proving that aesthetics are deeply tied to the narrative context of an image.
Conclusion
This work establishes a comprehensive roadmap for assessing color temperature aesthetics. By introducing the ICTAA task, the ICTA2Net framework, and the ICTAA240K dataset, we provide the tools necessary for more context-aware image assessment. Future work will expand this framework to include other color attributes like hue and saturation, further bridging the gap between machine and human aesthetic perception.
Summary updated