A Retinex-Based Local Tone Mapping Algorithm Using L0 Smoothing Filter - Advances in Image and Graphics Technologies - page 41

are thus necessary to better preserve the details and local contrast in images.
However, local tone mapping operators are at a higher cost, and may cause halo
artifacts because the local tone mapping operators consider the local information in
the mapping processing for each individual pixel.
It doesn't matter whether the scene is dark or bright, the Human Visual System
(HVS) can rapidly adapt to the luminance of the scene; therefore the HVS is the best
tone mapping operator [1]. It is important to indicate that the HVS consists of a global
and a local adaption. We can apply the HVS model to our tone mapping algorithm.
We have proposed a local tone mapping algorithm in this paper which reflects well
the HVS. Our method helps to solve the problems encountered when applying the
previously mentioned tone mapping algorithms, namely it does not produce halo
artifacts and provides good color rendition. The specifics of the proposed method will
be described in details below. In order to provide a better understanding, we will first
demonstrate how a global tone mapping operator is applied to the luminance image,
which imitates the initial adaption of the HVS; second, we will explain how to apply
the retinex-based local tone mapping operator to compress the contrast while
preserving the details; and finally, through the color correction, we will show the
achievement of the tone-mapped image.

2 Previous Works

2.1 Retinex Algorithm

The retinex theory, first proposed by Land [2], intends to explain how the human
visual system extracts reliable information from the world despite changes of
illumination. Many image processing experiments [3] show that the retinex theory is
consistent with lightness-color constancy theory. In other words, the color of objects
is independent with illumination, and it only relies on the surface reflectance of
objects.
In this paper, we will use retinex theory to solve the problem of high dynamic
range compression. According to the retinex theory, the given image

Ixy satisfies (, )

formula (1):

Ixy (, ) = Rxy (, ) × Lxy (, ) (1)

in which

Lxy is the illumination image. The
reflectance image R includes all the details of the original image, corresponding to the
high-frequency components, while the illumination image L corresponds to the
low-frequency components.
In order to compress the dynamic range of images, we should eliminate the effect
of uneven illumination. Therefore, for calculating detail image R , illumination image
L needs to be estimated. Then, subtracting illumination from the original image in the
log-domain, gets dynamic-range-compressed image. The retinex algorithm [3] is
given by

R xy is the reflectance image, and (, ) (, )

R xy (, ) log (, ) log( (, )*(, )) = Ixy − Fxy Ixy (2)