Adaptive Monte Carlo Retinex Method for Illumination and Reflectance
Separation and Color Image Enhancement
Alexander Wong, David A. Clausi, and Paul Fieguth
Vision and Image Processing Group
Department of Systems Design Engineering
University of Waterloo, Waterloo, Canada
{a28wong,dclausi,pfieguth}@uwaterloo.ca
Abstract
A novel stochastic Retinex method based on adaptive
Monte Carlo estimation is presented for the purpose of il-
lumination and reflectance separation and color image en-
hancement. A spatially-adaptive sampling scheme is em-
ployed to generate a set of random samples from the image
field. A Monte Carlo estimate of the illumination is com-
puted based on the Pearson Type VII error statistics of the
drawn samples. The proposed method takes advantage of
both local and global contrast information to provide bet-
ter separation of reflectance and illumination by reducing
the effects of strong shadows and other sharp illumination
changes on the estimation process, improving the preser-
vation of the original photographic tone, and avoiding the
amplification of noise in dark regions. Experimental results
using monochromatic face images under different illumina-
tion conditions and low-contrast chromatic images show the
effectiveness of the proposed method for illumination and
reflectance separation and color image enhancement when
compared to existing Retinex and color enhancement tech-
niques.
1. Introduction
An ongoing challenge in computer vision is alleviating
unwanted global and local illumination variations. In prac-
tical computer vision applications such as video surveil-
lance [1] and face recognition [2], images and videos
are often acquired in different unconstrained environments
where illumination can vary significantly within the ac-
quired scene. For example, lighting in outdoor environ-
ments can change significantly over the course of the day,
resulting in images of the same scene acquired at different
times of the day to appear very different from an image in-
tensity perspective. Similarly, the same objects can appear
very different due to differing or changing lighting condi-
tions in indoor environments. Such global and local illu-
mination variations make it difficult for computer vision al-
gorithms to recognize objects in a reliable and consistent
manner. In the realm of photography, obtaining images
with good contrast is desired, which is often not possible to
capture directly due to illumination variations in the scene.
Therefore, methods for alleviating the effects of global and
local illumination variations are sought.
One particularly effective class of approaches for reduc-
ing the effects of illumination variations is that based on
Retinex theory [3], where images are decomposed into their
individual illumination and reflectance components prior to
further processing. Then the reflectance information can be
used to achieve reliable object recognition that is invariant
to illumination conditions. Also, the illumination informa-
tion can then be modified independent of the reflectance in-
formation to achieve improved image contrast while avoid-
ing a washed-out appearance.
Retinex methods can be generally divided into two main
groups: i) global Retinex methods, and ii) local Retinex
methods. In global Retinex methods [4, 5, 6], pixel in-
tensity information along multiple random walks around
the image (with each walk ending at the pixel being esti-
mated) are used to estimate the reflectance of the image,
from which the illumination of the image can be subse-
quently estimated. The primary difference between global
Retinex methods is in the path geometry used. By exploit-
ing global information in the reflectance and illumination
separation process, global Retinex methods are able to bet-
ter preserve the original photographic tone of the image.
However, global Retinex methods tend to have poor detail
recovering, particularly in dark regions [7]. In local Retinex
methods [8, 9, 10, 11], the neighboring pixel intensities are
used to estimate the illumination of the image, from which
the reflectance of the image can be subsequently estimated.
The primary difference between local Retinex methods is
2009 Canadian Conference on Computer and Robot Vision
978-0-7695-3651-4/09 $25.00 © 2009 IEEE
DOI 10.1109/CRV.2009.24
108
Authorized licensed use limited to: University of Waterloo. Downloaded on December 20, 2009 at 19:34 from IEEE Xplore. Restrictions apply.