Variational Bayesian method for Retinex.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Image Processing

Background:

• Human color perception is complex and not fully understood.
• Existing Retinex methods often require additional information or lack robustness.
• Modeling the human visual system (HVS) for image interpretation is an ongoing challenge.

Purpose of the Study:

• To develop a variational Bayesian method for Retinex that simulates and interprets human color perception.
• To construct a hierarchical Bayesian model for reflectance and illumination estimation.
• To enable simultaneous estimation of unknown images and hyperparameters.

Main Methods:

• Utilized Gibbs and gamma distributions for prior distributions of reflectance, illumination, and model parameters.
• Defined energy functions using total variation and smooth functions based on piecewise continuous reflectance and spatially smooth illumination assumptions.
• Applied variational Bayes approximation to derive posterior distributions for parameter estimation.

Main Results:

• The proposed method demonstrates efficiency in estimating unknown image parameters and hyperparameters simultaneously.
• Achieved competitive performance without requiring additional information about unknown parameters.
• Outperformed non-Bayesian Retinex methods when prior information was incorporated.

Conclusions:

• The variational Bayesian Retinex method provides a robust framework for modeling human color perception.
• The method offers a powerful tool for image analysis and interpretation, particularly in low-light or challenging visual conditions.
• This approach enhances Retinex-based image processing by integrating Bayesian principles for improved accuracy and reduced reliance on external data.