Autoencoder-based training for multi-illuminant color constancy.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Color constancy is crucial for human vision, enabling object color perception under varying illumination.
• Replicating color constancy in software is challenging due to the ill-posed nature of image data (RGB values).
• Existing methods often struggle with multi-illuminant environments and rely on handcrafted or single-illuminant learned assumptions.

Purpose of the Study:

• To develop a robust method for achieving color constancy in images with multiple light sources.
• To address the limitations of current techniques in complex, real-world lighting conditions.
• To propose a deep learning approach for separating illumination and reflectance components.

Main Methods:

• Utilized an autoencoder architecture for image reconstruction.
• Trained the autoencoder to decompose images into distinct illumination and reflectance layers.
• Integrated illumination estimation with clustering for illumination segmentation.

Main Results:

• The proposed method successfully estimates illumination and reflectance in multi-illuminant scenes.
• The technique demonstrated superior performance compared to other tested methods in multi-illuminant scenarios.
• The approach proved invariant to the number of light sources present in the scene.

Conclusions:

• The autoencoder-based method offers a significant advancement in multi-illuminant color constancy.
• The technique provides a flexible solution, usable directly or with segmentation for illumination mapping.
• This work advances the capability of software to accurately perceive object colors under diverse lighting conditions.