Elongated Strip Oil Spill Segmentation Based on A
Cooperative Model
Mengmeng Di, Huajun Song, and Peng Ren
College of Information and Control Engineering,
China University of Petroleum,
Qingdao 266580, China.
Chunbo Luo
School of Computing,
University of The West of Scotland,
Paisley PA1 2BE, UK.
Email: [email protected]
Abstract—This paper describes a novel framework to segment
elongated strips of marine oil spill based on a cooperative model.
We construct a higher-order energy function which characterizes
data consistency and smoothness of varying orders in an inte-
grated framework. We model the data consistency in terms of a
finite Gamma mixture and then estimate its parameter values
using an expectation maximization (EM) algorithm. In order
to characterise the smoothness term we exploit a cooperative
model which not only penalizes the length but also addresses
the diversity of the object boundary. This advantage makes our
framework more effective for segmenting elongated strip oil spill
than traditional energy minimization methods such as graph cuts.
Our framework is further evaluated using practical SAR images
of oil spill recordings and its effectiveness is validated.
Keywords—Cooperative model, Oil spill segmentation, Graph
cuts.
I. INTRODUCTION
The global economy relies heavily on oil resources and
oil spill accidents from oil drilling platforms, ships or pipes
have occurred frequently. Oil spilled to marine environment
can cause severe harms to the natural environment and local
ecosystem and damage the aquaculture and economy [1].
Therefore, early detection of oil spill on marine surface is
particularly important. Of all the monitoring measures, Syn-
thetic Aperture Radar (SAR) installed on satellites have the
advantages of all-weather and all-time operation [2] which
make it a desirable means for detecting and monitoring oil
spilling [3].
In order to effectively detect and recognize oil spill on
ocean surface, the segmentation of oil spill areas in SAR
images is highly required and has been extensively studied
[4][5]. According to the principles of SAR imaging, the grey
values within one area of the SAR image may vary slightly
if the associated regions in the scene have similar reflection
characteristics and scattering properties. On the other hand,
different real world textures exhibit different grey values in
the SAR image. As a result, the grey level distribution of
a SAR image is rather complex and the oil spill image
segmentation based on such distribution becomes a difficult
task. In addition, weather conditions can change the marine
surface and reflection, and thus make oil spill segmentation be
more complicated.
In research literature, generic image segmentation methods
have been widely applied for oil spill segmentation [6]. One
straightforward way is to set a threshold for the grey values and
segment the SAR image into oil areas and non-oil areas [7][8].
This pixel-wise strategy is efficient in implementation, but
it neglects the pairwise relationship between the neighboring
pixels and may cause ambiguity if the noise level is high which
is usually the case for SAR imaging. Another category of
commonly used graph-cuts based methods combine pixel and
pairwise penalties together to yield more robust segmentation
schemes for processing generic images. However, SAR images
with oil spills are different from generic images for the fact
that there are a considerable amount of oil spill areas exposed
with the shape of elongated strips. In this case, graph cuts
[9] based methods tend to be much ineffective because the
pairwise terms in graph cuts encourage smooth segmentations
by penalizing the assignment of different labels to neighbor-
ing pixels, therefore, the graph cuts cannot characterize the
boundaries of elongated strips properly.
To address the drawbacks of the existing methods, we
exploit a cooperative model [10] which has been recently
proposed by Pushmeet Kohli, to model the higher-order terms
of the energy function for oil spill segmentation. Instead of
favouring short oil boundaries by penalizing the number of
label discontinuities, our method focuses on the congruous
boundaries by penalizing the number of types of label dis-
continuities and thus has the advantages of preserving the
boundaries of elongated strips and improved performance. The
experiment results confirm that our method is more effective
in segmenting the elongated stripped oil spills of SAR images
than existing methods such as thresholding and graph cuts.
II. ENERGY FORMULATION
We formulate the problem of oil spill segmentation in terms
of an energy minimization scheme. The first step is to construct
the energy function for segmentation. We use the set P =
{p
1
, p
2
, ..., p
N
} to denote the pixel set of a SAR image, and I
to denote the pixel gray value in set P . The object of oil spill
segmentation is to assign each pixel p
i
∈ P a binary value
x
i
∈ {0, 1}, where the label 0 represents an oil filling area
and the label 1 represents a background area. The set of all
binary variables x
i
for p
i
is denoted by X.
资源评论
weixin_38713057
- 粉丝: 3
最新资源
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
