Edge-moment-based color constancy using illumination-coherent regularized regression.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Color constancy is crucial for accurate image interpretation under varying illumination.
• Existing methods often struggle to maintain image similarity with corresponding illuminants.
• A gap exists in literature regarding illumination-coherent image similarity for color constancy.

Purpose of the Study:

• To propose a novel edge-moment-based algorithm for color constancy.
• To enhance image similarity by incorporating illuminant information in a coherent space.
• To achieve state-of-the-art performance in color constancy tasks.

Main Methods:

• Utilizing color edge moments for scene image representation.
• Employing Canonical Correlation Analysis (CCA) to project features into an illumination-coherent space.
• Applying a Mixture of Gaussians (MoG) model to create consistent subspaces.
• Implementing iterative L2-norm regularized regression within subspaces to learn illuminant correlations.
• Developing a soft fusion strategy for subspace estimations based on posterior probabilities.

Main Results:

• The proposed algorithm demonstrates superior performance compared to existing state-of-the-art methods.
• Extensive experiments on standard datasets validate the effectiveness of the approach.
• Both intra- and inter-dataset evaluations confirm the algorithm's robustness and accuracy.

Conclusions:

• The developed edge-moment-based algorithm effectively achieves color constancy by leveraging illumination-coherent image similarity.
• The divide-and-conquer strategy within consistent subspaces enhances the robustness of illuminant estimation.
• This work establishes a new benchmark for color constancy algorithms in computer vision.