Modeling surface color discrimination under different lighting environments using image chromatic statistics and...

Area of Science:

• Computer vision
• Color science
• Human perception

Background:

• Understanding human color perception is crucial for applications like image processing and digital displays.
• Existing models based on chromatic statistics struggle to capture the complexity of human color discrimination across diverse lighting conditions.

Purpose of the Study:

• To model human discrimination thresholds for object colors under various lighting environments.
• To compare the predictive power of chromatic statistics models versus machine learning approaches.
• To identify key visual features influencing color perception.

Main Methods:

• Developed and tested 60 distinct chromatic statistics models.
• Trained convolutional neural networks (CNNs) on a dataset of 160,280 images, using both ground-truth and human-response labels.
• Performed region-of-interest analysis on the CNNs to understand feature importance.

Main Results:

• No single chromatic statistics model adequately predicted human color discrimination thresholds across all tested conditions.
• CNNs trained on human responses achieved near-perfect prediction of human thresholds.
• Modifying chromatic models to focus on lower object regions significantly enhanced their predictive performance.

Conclusions:

• Convolutional neural networks offer a powerful approach for modeling human color discrimination.
• Human responses are critical for training effective predictive models.
• The spatial characteristics of object regions play a significant role in color perception under different lighting.