Segmentation and estimation of spatially varying illumination

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Accurate illumination estimation is crucial for image analysis and color correction.
• Existing methods often require multiple images or manual intervention.
• Segmenting multiple light sources from a single image remains a challenge.

Purpose of the Study:

• To develop an unsupervised method for segmenting illuminant regions and estimating the illumination power spectrum from a single image.
• To address the challenge of analyzing scenes lit by multiple light sources.
• To provide a robust solution for illumination estimation without prior knowledge.

Main Methods:

• Probabilistic clustering in image spectral radiance space for illuminant region segmentation.
• Optimization framework maximizing likelihood with spatial smoothness constraints.
• Coordinate-ascent optimization for mixture model weights, pixel sets, and posterior probabilities.
• Estimation of per-pixel illuminant power spectrum using posterior probabilities.

Main Results:

• Successfully segmented illuminant regions and estimated illumination power spectrum.
• Demonstrated effectiveness on hyperspectral and trichromatic image datasets.
• Outperformed alternative methods in illumination region segmentation, color estimation, and color correction.

Conclusions:

• The proposed unsupervised method offers an effective solution for illuminant region segmentation and power spectrum estimation.
• The approach is robust and applicable to various image types and lighting conditions.
• This work advances the field of single-image illumination analysis and color constancy.