AN EFFECTIVE FALSE-ALARM REMOVAL METHOD
BASED ON OC-SVM FOR SAR SHIP DETECTION
Xiaoting Yang
1
, Fukun Bi
2
, Ying Yu
3
, Liang Chen
1,
*
1
Beijing Key Laboratory of Embedded Real-time Information Processing Technology, Beijing
Institute of Technology,Beijing 100081,China
2
School of Electronic Information Engineering, North China University of Technology, Beijing
100144, China
3
School of Information Science and Engineering, Yunnan University, Kunming 650091, China
chenl@bit.edu.cn
Keywords: OC-SVM, ship detection, false alarms
removal, parameters analysis, performance comparison.
Abstract
Automatic ship detection from SAR remote sensing
imagery is very important, with a wide array of
applications in areas such as national marine safety,
vessel traffic services, and naval warfare. However, SAR
remote sensing images with the high inhomogeneity of
sea clutter and complex scene result in a large number of
false alarms. This paper focuses on the problem of false
alarm removal and coverts it as a classification problem
with respect to ship targets and false alarms. However, in
real application: 1) Compared to false alarms the number
of ships is much smaller, and 2) false alarms with diverse
categories and no uniform characteristics are difficult to
be distinguished from ships. To solve the problems above,
we present a new method based on one-class SVM (OC-
SVM) combined with grid optimization that only use the
data of ships to train classifier in the absence of false
alarms. First, 2P-CFAR is adopted to extract ship
candidates and 14 features mainly focusing on ship targets
are selected. Second, grid optimization is exploited to
obtain parameters for OC-SVM training, where it is a
compromise between the probability of false alarm and
detection rate. Experimental results of OC-SVM on a
large SAR image set demonstrate that in comparison with
other state-of-the-art methods like SVM, BP neural
network, our approach can achieve higher detection
accuracy.
1 Introduction
Synthetic aperture radar (SAR) with capacity to image
day and night under most meteorological conditions
became the state-of-the-art technique for ship detection [1]
and play a prominent role in the national marine safety,
ocean management, and monitoring illegal fishing [2][3].
Numerous studies have been performed. CFAR [4]
detection, which is famous for its constant false alarm
probability and adaptive threshold has been used widely,
such as Gamma, Weibull, K, etc.[3,5]. And Gui Gao [6]
proposed Parzen-Window-Kernel-Based CFAR algorithm
for ship detection. However, these methods all lead to a
high false alarm rate due to SAR remote sensing images
with the high inhomogeneity of sea clutter and cannot
remove false alarms effectively. The removal of false
alarm is still a challenge of today.
In fact, to distinguish ships from false alarms can regard
as a classification problem. And various classifiers have
been used in numerous fields, such as K-means cluster [7],
Support Vector Machine (SVM) [8] and One-Class SVM
[9].
K-means clustering is a simple method, no need to
estimate the distribution of samples, but it is unsuitable
here because the effects of noises and the diversity among
false alarms make it difficult to select the appropriate K
and the cluster centers, which are directly related to the
performance of the final classification. BP neural network
[10] is a feed-forward network, and is one of the most
widely used neural networks, with its self-learning, self-
organizing, and adaptive capabilities. However, there
may be over-fitting and lower generalization problems. In
the process of ship detection due to the small training sets,
which can not guarantee the training samples include all
types of ships and false alarms, so that the classifier can
not correctly predict unknown samples. Support Vector
Machine (SVM) [8] is originally for dichotomous
questions, the main idea is to map the data set to a high
dimensional feature space, and in this feature space to
find an optimal hyperplane to separate the two types of
samples. However, in the process of ship detection not
only false alarms with diverse categories, but they also
have various forms of internal features, resulting in a
simple hyperplane can not achieve the goal to classify
ships and false alarms effectively. Even though
constructing multiple classifiers, due to the uncertainty
and diversity of false alarms, we also hard to determine
the number of classifiers and extract optimal features.
What’s more, the number of ship is much smaller than
false alarms, which lead to the deviation of hyperplane.
Aiming at the problems of unbalanced number between
ship targets and false alarms, also the diverse categories
and no uniform characteristics for false alarms to be
distinguished from ships, We present a new method based
on one-class SVM (OC-SVM) to escape false alarms for
ship detection, mainly focusing on the ship targets to
extract features and train classifiers. On the one hand this