Deep Dichromatic Guided Learning for Illuminant Estimation.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Traditional dichromatic reflection models struggle with specular separation, limiting real-world applications.
• Deep neural network (DNN) methods enhance illuminant color estimation but lack explainability.

Purpose of the Study:

• To propose a novel DNN architecture for accurate and explainable dichromatic illuminant estimation.
• To address limitations in separating specular reflections and understanding DNN performance.

Main Methods:

• Developed a DNN architecture to learn dichromatic planes and their confidences using a novel loss function.
• Estimated illuminant color via a weighted least mean square of learned planes.
• Introduced dichromatic guided learning for enhanced analysis.

Main Results:

• Achieved state-of-the-art results in benchmark evaluations for color constancy.
• Generated a map detailing color and regional contributions to illuminant estimation.
• Enabled in-depth analysis of illuminant estimation success and failure cases.

Conclusions:

• The proposed method offers compelling performance in color constancy.
• Dichromatic guided learning provides valuable insights into illuminant estimation processes.
• The approach enhances both accuracy and interpretability in real-world image analysis.