Fast Image Segmentation With
v
L
Distance Between Local Cumulative Histograms
S.Q Zhang
a
, W.L Wang
b
, J.F Lu
c
, L Li
d
Wenlei Wang
b
Institute of Graphics and Image
Computer Science and Technology HDU
Hangzhou, China 310018
e-mail: 312037141@qq.com
Abstract
ü
This paper presents two new region-based active
contour models in a variation framework for image
segmentation. In these models, to quantify the similarity
measurement between two regions, we propose to compare
their respective cumulative histograms with the
1, 2
,LL L
∞
distances which are often applied in image comparison. For
example, to compute the similarity, the Kolmogorov-
Smirnov test by using the
L
∞
distance is the most popular
one. Our first energy model is based on minimizing
v
L
distance between two cumulative histograms of different
regions, together with a geometric regulation term that
penalizes the length of region boundaries. The second model
is the localization of the first model with the popular local
method. To solve these models, we describe a fast
minimization algorithm with the Split-Bregman method,
which can find the global minimizer. Finally, to illustrate the
robust and accurate segmentations with the proposed
models, we show some experimental results on some images.
Keywords: Active contours, Image segmentation,
Cumulative histogram, Split-Bregman Method
I.
I
NTRODUCTION
Active contour models have been extensively applied
to image segmentation. Existing active contour models
can be categorized into two major classes: edge-based
models, and region-based models. Edge-based models
utilize boundary information to stop the evolving contours
on the object boundaries. This type of highly localized
image information is adequate in some situations, but has
been found to be very sensitive to image noise and highly
dependent on initial curve placement. Region-based active
contours are robust to noise. There are many advantages
of region-based approaches when compared to edge-based
methods including robustness against initial curve
placement and insensitivity to image noise. Furthermore,
they are able to detect objects with either sharp or smooth
edges. However, popular region-based active contour
models[1] tend to rely on intensity homogeneity in each
of the regions to be segmented. For example, the Chan-
Vese (CV)models are based on the assumption that image
intensities are statistically homogeneous in each region. In
fact, intensity inhomogeneity often occurs in medicine
images. Intensity inhomogeneity can be addressed by
more easy models , such as Vese and Chan and Tsai et al.
independently proposed two similar region-based models
for more general images. Aiming at minimizing the
Mumford-Shah functional[2], both models cast image
segmentation as a problem of finding an optimal
approximation of the original image by a piecewise
smooth function. To accurately segment these objects, a
new class of active contour energies is presented, which
utilizes local information, but also incorporates the
benefits of region-based techniques, for example local
binary fitting (LBF)model[3]. Whatever using global
statistics information or using local statistics information,
we will find an (local) optimal approximation of the
(local) original image according to the foreground. In
order to design a data term in the model, we have to
compare the optimal image with the original image.
Image comparison is one of the essential tasks of image
retrieval. The image histogram is a frequently used
technique of similarity measurement of some images.
However, except the histogram, one can compute cally
the cumulative distribution function (or the cumulative
histograms) which has an advantage of application of
nonparametric statistical tests, such as Kolmogorov-
Smirnov test. Other nonparametric tests for distribution
comparison are available, such as Friedman or Wilcoxon
test[4] as described in the reference book. However, the
Kolmogorov-Smirnov test is the most popular, perhaps
due to simplicity of computations. Several authors used
Kolmogorov-Smirnov test for the image segmentation.
See[5], the
L
∞
distance between the two cumulative
distribution function is used to define a similarity measure
of two cumulative distribution function , Of course, the
1
L
and
2
L
distance can be also used to define a similarity
measure of two cumulative distribution function. In [6],
Tony Chan et al. presented a segmentation model based
on minimizing the Wasserstein distance with exponent 1,
which is able to fairly compare two histograms. In fact,
the Wasserstein distance with exponent 1 is
1
L
the
distance between two histograms.
The layout of the paper is as follows. Section 2
presents facts from Kolmogorov-Smirnov test; Section 3
gives a new segmentation model based on Kolmogorov-
Smirnov test; Section 4 shows the fast algorithms and
discretization for solving the proposed model. Section V
illustrates some experimental results and comparison with
other methods in some images.
II.
L-
METRIC AND
K
OLMOGOROV
-S
MIRNOV TEST
Let
be a stochastic variable with density
function
()fx
, the cumulative density (distribution)
is
defined in terms of the probability
P
by:
2016 Nicograph International
978-1-5090-2305-9/16 $31.00 © 2016 IEEE
DOI 10.1109/NicoInt.2016.19
100