Improving color constancy by selecting suitable set of training images

Area of Science:

• Computer Vision
• Image Processing

Background:

• Regression-based color constancy (CC) methods offer competitive performance.
• Current methods often apply a single correction matrix to all images, which is suboptimal due to varying image conditions.

Purpose of the Study:

• To improve regression-based CC by addressing the limitation of using a single correction matrix.
• To develop a strategy for automatically selecting optimal training data for each test image.

Main Methods:

• Mathematical analysis of factors influencing regression-based CC performance.
• Design of principled rules for automatic selection of training images.
• Learning an optimal correction matrix tailored to each test image.

Main Results:

• Significant performance improvements in regression-based CC (e.g., Corrected Moment) on benchmark datasets.
• Demonstration of increased accuracy by adapting correction matrices to individual images.
• Validation of the proposed strategy's effectiveness on multiple regression-based CC approaches.

Conclusions:

• Automatic training data selection is a key strategy for enhancing regression-based CC.
• The proposed method offers a more robust and accurate approach to color constancy.
• The strategy is broadly applicable to various regression-based CC algorithms.