Self-Supervised Color-Concept Association via Image Colorization.

Area of Science:

• Computer Vision
• Data Visualization
• Human-Computer Interaction

Background:

• Effective color-concept associations in visualizations enhance semantic interpretation.
• Manual creation of color-concept datasets is time-consuming and expensive.
• Existing methods rely on supervised learning and human-annotated data.

Purpose of the Study:

• To develop an automated method for extracting color-concept associations from concept images.
• To overcome the limitations of manual data collection and supervised learning approaches.
• To enable the creation of semantically meaningful color palettes for data visualizations.

Main Methods:

• A self-supervised learning approach utilizing automatic image colorization.
• Fine-tuning a pre-trained colorization model (ImageNet) on concept-specific image datasets.
• Predicting pixel-wise color probability distributions in the Lab color space and converting them into color ratings.

Main Results:

• The self-supervised method effectively extracts color-concept associations without ground truth data.
• Achieved competitive results compared to supervised methods, particularly in metrics like Earth Mover's Distance (EMD) and Entropy Difference (ED).
• The proposed method's results align well with human perception of color distributions.

Conclusions:

• The developed self-supervised colorization method provides an efficient alternative for generating color-concept associations.
• This approach has the potential to significantly reduce the cost and effort associated with creating semantic color palettes.
• The method offers a promising direction for improving the interpretability of data visualizations through better color mapping.