Decision fusion of sparse representation and support vector machine for SAR
image target recognition
Haicang Liu, Shutao Li
n
College of Electrical and Information Engineering, Hunan University, Changsha, Hunan 410082, China
article info
Article history:
Received 7 April 2012
Received in revised form
7 December 2012
Accepted 3 January 2013
Communicated by: Shiguang Shan
Available online 5 March 2013
Keywords:
Sparse representation
SVM
Decision fusion
SAR
Target recognition
abstract
We propose a decision fusion method of Sparse Repr esentation (SR) and Support Vector Machine (SVM)
for Synthetic Aperture Radar (SAR) image target recognition in this paper. First, a fast SR classifier (FSR-
C) with Matching Pursuit (MP) solution is proposed. In the FSR-C, the dictionary is compose d of training
images. Just one nonzero element in SR coefficient of the testing image is found out based on MP, and
the testing image is classified through the location of the nonzero element. To further improve the
recognition accuracy, the SVM classifier (SVM-C) is selected. In SVM-C, PCA feature is extracted, and for
seeking the linear separating hyperplane, the RBF kernel function is used in mapping the training
vectors into high dimensional space. The results of the FSR-C and the SVM-C are fused obeying Bayesian
rule to make the decision. The Moving and Stationary Target Acquisition and Recognition (MSTAR) SAR
image database is used to test the performance of the proposed method. The experimental results show
that the FSR-C can predict testing SAR images with considerable recognition accuracy and high real-
time ability, and the decision fusion recognition method can improve the recognition accuracy and still
be fast.
& 2013 Elsevier B.V. All rights reserved.
1. Introduction
Taking advantage of acquiring image in inclement weather or
during night as well as day, Synthetic Aperture Radar (SAR) image
is widely applied in civilian and military. SAR image target
recognition is the most important step in SAR image interpreta-
tion and analysis [1].
Target recognition generally consists of three processes, i.e.,
preprocessing, feature extraction and classification [2]. The image
feature is extracted, and it directly affects the probability of
correct classification (PCC) [3]. There are many methods for SAR
image target feature extraction, such as Principal Component
Analysis (PCA), Independent Component Analysis (ICA), Linear
Discriminant Analysis (LDA) and Discrete Wavelet Transform
(DWT) [4–6]. These methods have different characteristics, and
fit different tasks. A variety of target classification algorithms
have been used in SAR image, such as Support Vector Machine
(SVM) [7,8], Adaboost [9], Neural Network (NN) [10], Gaussian
Mixture Models (GMM) [11,12] and so on.
Wu Tao et al. [8] studied SAR image target recognition with
SVM classifier. They divided the training samples into several
groups according to the SAR image aspect angles and train SVM
classifier for each group. In the test, they first measured the aspect
angle through segmentation, and then extracted PCA feature. At
last, the result is predicted in the corresponding SVM classifier.
The method got high recognition accuracy. However, in this
method, the PCC would decrease with the increasing of the aspect
angle interval, and target segmentation is very time-consuming
and hard to achieve high accuracy.
Sparse Representation (SR) is a new method for target recog-
nition [13,14]. Wright et al. [13] proposed the SR method for face
recognition, and Thiagarajan et al. [14] modified the approach for
classifying targets in SAR image. In Ref. [13] or [14], the dictionary
is composed of the training vectors that are normalized, and the
sparse representation of the testing data is computed with the
dictionary. The sparse representation is a locally linear approx-
imation with respect to the corresponding class. The testing
image is reconstructed from the sparse representation coefficient
and the dictionary, and is used to classify. The Orthogonal
Matching Pursuit (OMP) is used in Ref. [14] instead of l
1
norm
in Ref. [13] to solve the sparse representation of the testing image.
The SR method with OMP solution (SR-OMP) proposed by
Thiagarajan et al. is still time-consuming because of the iteration
in OMP and the reconstruction process of the testing image.
To improve the target recognition performance, many fusion
methods are proposed [15–18]. Fig. 1 displays three types of
decision fusion methods that are often used for SAR image
recognition. Fig. 1 (a) is a multi-angle decision fusion method
Contents lists available at SciVerse ScienceDirect
journal homepage: www.elsevier.com/locate/neucom
Neurocomputing
0925-2312/$ - see front matter & 2013 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.neucom.2013.01.033
n
Corresponding author. Tel.: þ86 73 188 822 866.
E-mail addresses: shutao_li@yahoo.com.cn, shutao_li@hnu.edu.cn (S. Li).
Neurocomputing 113 (2013) 97–104