*s.greg.jones@narelab.com; phone 1 222 555-1234; fax 1 222 555-876; narelab.com
A Robust Active Contour Edge Detection Algorithm Based on Local
Gaussian Statistical Model for Oil Slick Remote Sensing Image
Yu Jing*
a
, Yaxuan Wang
a
, Jianxin Liu
a
, Zhaoxia Liu
a
a
School of Software, Dalian University of Foreign Languages, Dalian116044, P.R.China
ABSTRACT
Edge detection is a crucial method for the location and quantity estimation of oil slick when oil spills on the sea. In this
paper, we present a robust active contour edge detection algorithm for oil spill remote sensing images. In the proposed
algorithm, we define a local Gaussian data fitting energy term with spatially varying means and variances, and this data
fitting energy term is introduced into a global minimization active contour (GMAC) framework. The energy function
minimization is achieved fast by a dual formulation of the weighted total variation norm. The proposed algorithm avoids
the existence of local minima, does not require the definition of initial contour, and is robust to weak boundaries, high
noise and severe intensity inhomogeneity exiting in oil slick remote sensing images. Furthermore, the edge detection of
oil slick and the correction of intensity inhomogeneity are simultaneously achieved via the proposed algorithm. The
experiment results have shown that a superior performance of proposed algorithm over state-of-the-art edge detection
algorithms. In addition, the proposed algorithm can also deal with the special images with the object and background of
the same intensity means but different variances.
Keywords: Edge Detection, Oil Spill Remote Sensing Image, Gaussian Distribution Fitting, Global Minimization Active
Contour Model, Intensity Inhomogeneity, Dual Formulation
1. INTRODUCTION
Recently, oil spill accidents and illegal oily discharges from tank cleaning or bilge pumping occur frequently on the sea,
and these situations represent a serious threat to the marine environment and cause great losses of energy sources. Early
detection, monitoring, containment, and cleanup of oil spill are crucial for the protection of the environment. The uses of
satellite-based synthetic aperture radar (SAR) images and aircraft-based infrared (IR) aerial images are the cost-effective
ways to monitor large oil spill areas and verify the oil spill
[1, 2, 3]
. The technology of edge detection is an important tool
for the location and acreage calculation of oil slick on the sea by aerial remote sensing. Whenever we need to identify oil
spill, confirm the location, or get the shape and acreage of oil spill, we have to get the edge information of oil slick
images first
[4, 5, 6]
.
Oil slick remote sensing images generally have the following characteristics. (1) Generally dark, edge quite blurry. (2)
Gradient difference between oil and seawater is small and intensity inhomogeneity often occurs. (3) Containing massive
stripe noise and speckle noise. (4) No fixed shape
[7]
. Because of these characteristics, edge detection of oil slick images
is a difficult and complex task.
Recently, active contour models (ACMs) have been extensively applied to image processing fields, such as image edge
detection, image denoising, and image restoration. Edge detection based on ACMs is one of the solutions of image
segmentation. Compared with the conventional edge detection methods, the ACMs applied to edge extraction have the
following significant advantages: 1) The ACMs can achieve subpixel accuracy of object boundaries; 2) the ACMs can be
easily formulated under a principled energy minimization framework and allow the incorporation of various prior
knowledge, such as shape and intensity distribution, and moreover, the ACMs have strong robustness for image noise
and edge break; and 3) the ACMs can provide smooth and closed contours, thus avoiding the process of postprocessing
in the traditional edge detection methods
[8]
. Generally speaking, the existing ACM methods can be classified into two
types: edge based models and region-based models. The edge-based models use local edge information to attract the
active contour toward the object boundaries. The region-based models utilize the image statistical information to guide
the motion of the active contour and have more advantages over the edge-based models
[10]
. Nowadays, most of the
popular region-based models are the Chan and Vese model
[11]
and the Li’s models
[9, 12, 13]
. However, these models have
some common drawbacks. In particular, the existence of local minima in the process of the active contour energy
minimization makes the initial guess critical to get satisfactory results and time consuming. In order to accurately extract
edited by Xuping Zhang, David Erickson, Xudong Fan, Zhongping Chen, Proc. of SPIE Vol. 9620,
Proc. of SPIE Vol. 9620 962019-1