Improving RGB illuminant estimation exploiting spectral average radiance.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Color constancy aims to render images consistently under varying illumination conditions.
• Accurate illuminant estimation is crucial for effective color correction.
• Existing methods often struggle to effectively combine spectral and RGB color information.

Purpose of the Study:

• To develop an enhanced RGB color constancy method using neural networks and k-means clustering.
• To investigate the optimal integration of multispectral and RGB color data for illuminant estimation.
• To determine the most effective domain (spectral or RGB) for illuminant prediction and loss function definition.

Main Methods:

• A hybrid approach combining neural networks and k-means clustering for illuminant estimation.
• Exploration of different spatial resolutions for sampling RGB and spectral image data.
• Comparative analysis of predicting illuminants in spectral versus RGB domains.
• Evaluation of loss functions defined in either the RGB color or spectral domain.

Main Results:

• The best performance was achieved by predicting the illuminant in the spectral domain with an RGB color loss function.
• This method significantly improved the recovery angular error metric by 66% over spectral methods and 41% over RGB methods.
• The study identified optimal spatial resolutions for input data sampling.

Conclusions:

• Combining multispectral and RGB information with a neural network and k-means clustering yields superior color constancy.
• Predicting illuminants in the spectral domain with an RGB loss function is the most effective strategy.
• The proposed method represents a significant advancement in the field of color constancy.