65
2019 IEEE 4th International Conference on Signal and Image Processing
Spotlight SAR Image Recognition Based on Dual-Channel Feature Map
Convolutional Neural Network
Junjie Liu, Xiongjun Fu*, Kaiqiang Liu, Miao Wang, Chengyan Zhang, Qinning Su
School of Information and Electronics
Beijing Institute of Technology
Beijing, China
e-mail: fuxiongjun@bit.edu.cn
Abstract—Synthetic Aperture Radar (SAR) is widely used in
agriculture, remote sensing and many other fields due to its all-
weather working mode and its excellent penetration. However,
the decipherment of synthetic aperture radar imaging is very
difficult compared to optical images. This problem is even
worse in the SAR target recognition. Although the traditional
feature engineering method is helpful for SAR image
information content extraction, the effect is not satisfied with
the requirements in practice. Convolutional neural network is
an effective method to extract synthetic aperture radar
imaging features and recognize targets. In this paper, a dual-
channel feature map convolutional neural network (DCFM-
CNN) is proposed, using two different down sampling methods,
---- pooling and convolution, to extract features for SAR image
automatic target recognition (SAR-ATR). An average
recognition accuracy of 99.45% was achieved on MSTAR
public data set. Preprocessing of synthetic aperture radar
imaging is not needed here, and the target recognition is
completed by the CNN model. The proposed object recognition
approach is effective with low overhead.
Keywords-synthetic aperture radar imaging; deep learning;
dual-channel feature map; convolutional neural network; object
recognition
I. INTRODUCTION
Synthetic Aperture Radar plays an important role in
agriculture, forestry and other fields, and also it is
irreplaceable in the military field. First, the high-resolution
SAR system can obtain clear images of the target all time,
which is beneficial to identify the object of interest. Secondly,
the SAR system has penetrating power and can effectively
detect the hidden object. However, compared with optical
images, it is difficult to manually extract information of
interest in SAR images. Even if a suspected object is found,
it is difficult to identify the specific information of the object.
Convolutional neural networks automatically extract image
features through convolutional layers, omitting the process of
visually identifying SAR image objects, effectively solving
this problem. Using the convolutional neural network to
complete the SAR-ATR work, only need to know the type of
the object, the task of the feature engineering is all handed to
the computer for processing, and the neural network
extraction feature will be more complete, and is not limited
by the existing knowledge. Therefore, the extracted features
will be more sufficient, including features recognized by
existing knowledge, as well as features that are not
recognized by existing knowledge. These more abundant
features can improve the accuracy of object recognition.
To complete the SAR-ATR task, data preprocessing of
SAR image data is usually performed first; then the
preprocessed image data is classified using a classifier, and
the design of the classifier includes traditional machine
learning algorithms and neural networks. Through different
feature extraction methods and selection of classifiers, better
recognition results can be obtained. Commonly recognition
methods include artificial feature extraction combined with
traditional machine learning algorithms such as SVM and
KNN [1], and neural network methods [2-5]. In [2], a two-
channel deep convolutional neural network model is
proposed, which uses different pooling methods to extract
different feature joint predictions. In [3], a full convolutional
neural network is proposed, which uses convolutional layers
instead of pooling layers and fully connected layers to
complete the down sampling operation and recognition
function, which reduces the number of parameter values
introduced by the fully connected layers, improving the
efficiency of the network. In [4], a method using a neural
network without a fully connected layer to extract image
features and using SVM to classify the targets is proposed. In
[5], a multi-scale convolutional neural network and SVM
target joint prediction method is proposed. Different levels of
features are extracted from the network, so that low-level
features can be fully utilized. In addition, articles [6-9] use
other methods and have a good effect on the SAR-ATR
problem. Most of the above methods use a combination of
algorithms to achieve object recognition, which is more
complicated. However, the method in this paper only needs
to train a convolutional neural network model to recognize
object efficiently. It is simple to apply and has high accuracy.
This paper mainly studies the application of
convolutional neural network in SAR image target
recognition. An end-to-end dual-channel feature map
convolutional neural network (DCFM-CNN) is proposed for
MSTAR dataset. Only need to input SAR images of which
the type needs to be identified, the category of the target in
the SAR image can be returned. Compared with common
CNN structure, the structure of DCFM-CNN uses two
branches to extract features in different ways. One of the
branches performs normal convolution and pooling
operations, and the other branch is down sampled by
978-1-7281-3660-8/19/$31.00 ©2019 IEEE