Properties and performance of a center/surround retinex

Area of Science:

• Computer Vision
• Image Processing
• Computational Neuroscience

Background:

• Land's retinex model (1986) provides a theoretical framework for human vision's lightness and color constancy.
• Previous research has focused on the mathematical underpinnings of the retinex model, with limited experimental validation in image processing.

Purpose of the Study:

• To implement and experimentally test the latest version of Land's retinex model for image processing applications.
• To optimize retinex parameters for practical image rendition and dynamic range compression, independent of human perception modeling.

Main Methods:

• Implementation of Land's retinex model with adjustable parameters for surround space constant, logarithmic function placement, and gain/offset application.
• Evaluation of various functional forms for the retinex surround, including Gaussian and inverse square.
• Analysis of image rendition quality under different parameter settings and investigation of model performance on images violating gray world assumptions.

Main Results:

• The placement of the logarithmic function after surround formation significantly improves results.
• A Gaussian surround form outperforms the inverse square form suggested by Land.
• Optimal rendition is achieved with a canonical gain/offset applied post-retinex operation.
• The trade-off between rendition quality and dynamic range compression is governed by the surround space constant.

Conclusions:

• The implemented retinex model offers a practical approach to image processing, achieving effective dynamic range compression and rendition.
• Parameter optimization, particularly the placement of the logarithmic function and the choice of surround form, is crucial for retinex performance.
• Understanding the limitations of the retinex model, especially concerning gray world assumption violations, is key for its effective application in diverse image scenarios.