IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 29

Oveis et al.
scene SAR images by providing bounding boxes. In fact,
they have randomly placed some targets from MSTAR
dataset [8] in an image of fields, trees, and bushes to make
up complex large scene SAR image. They used a fast sliding
method to decompose the scene image into subimages
to avoid dividing targets into different subimages.
An interesting topic of object detection in SAR images
is ship detection. Kang et al. [134] proposed an R-CNN
for ship detection that is based on contextual information
and multilayer features. Zhao et al. [142], integrated the
visual attention model in frequency domain into a CNNbased
SAR ship detection framework to reduce the missed
detections in small and densely clustered areas. Bentes et
al. [143] proposed a framework in which ships in TerraSAR-X
images are detected by a standard CFAR algorithm
and then classified by a CNN. For ship classification
in SAR images, Wang et al. [144] studied transfer learning
and fine-tuning of VGG16, VGG19, Xception [38], and
InceptionV3 [145]. Lin et al. [146] exploited Squeeze and
Excitation mechanism to improve the performance of
Faster-RCNN for ship detection in SAR images. SENet
[147] provides a weighting mechanism for each channel
of the feature maps. Zhang et al. [148] proposed a gridbased
CNN architecture that is inspired by YOLO [137] to
decrease the computational time of ship detection problem
in SAR images. Ai et al. [149] proposed a feature fusion
scheme from two subnetworks for ship detection. They
utilized the low-level Haar-like features [150] and highlevel
deep features of a CNN subnetwork. They showed
that VGG-Net and Res-Net capture the high-level deep
features of the ship targets. However, these networks
ignore the local edge and texture information so their performance
is poorer than the multi-level feature fusionbased
discriminator of [149]. These low-level Haar-like
features include horizontal and vertical edges, lines with
different scales and rotations. Deng et al. [151] proposed a
feature reuse strategy for their CNN-based ship detection.
They used OpenSARShip dataset [70] that contains different
types of ships collected from the Sentinel-1 satellite.
SSDD has also been used in many ship detection studies.
This dataset, which consists of 2358 ships in 1160 SAR
satellite images, is labeled by Li et al. [71] by using LabelImg
[152] as an image annotation software [153]. In
[154]-[156], SAR images from the Chinese satellite Gaofen-3
that was launched in 2016 [157] are employed for
the proposed CNN-based ship detection architectures.
Other CNN-based SAR ship detection methods have been
discussed in [158]-[160].
RELATED TOPICS TO CNN AND SAR
In this section, we address different aspects regarding the
use of CNNs for SAR data processing. Different concepts
such as CV architectures, data augmentation, and transfer
MAY 2022
learning will be discussed. Moreover, the use of convolutions
in the generative adversarial network (GAN) and
sparse coding will be addressed.
COMPLEX-VALUED CNN
Some researchers have attempted to exploit the phase
inside CNNs because it is a well-known property of SAR
that the phase contains very important information. One
simple approach to transfer the phase information, which
is embedded in complex SAR images, into the conventional
RV-CNNs is to employ two channels where magnitude
and phase (or real and imaginary parts) are taken as
input. However, this is not an efficient approach and it is
better to extract features in CV domain to maintain the
integrality of information [161]. Mathematical formulations
for CV-CNN have been thoroughly addressed in
[162] and [163]. In CV-CNNs, all elements of CNNs
including input - output layers, convolution layers, activation
functions, and pooling layers are represented in complex
form. Moreover, the complex backpropagation
algorithm based on stochastic gradient descent has been
studied. Zhang et al. [164] proposed a CV-CNN for polarimetric
SAR image classification, where each pixel is classified
into known terrain types, such as forest, grass,
water, urban area, and sand. Their suggested network
takes a tensor with six channels as input data similar to
[80]. However, only the upper triangular part of the 33
PolSAR complex coherency matrix is used in [164]. Gao
et al. [165] proposed a CV-CNN for ISAR image formation.
They substituted conventional sigmoid function used
in [164] by extending ReLU to its CV version. The input
and output of this CV-CNN are radar echoes and expected
images, respectively. They used the classical turntable
model for generation of the training data. Sunaga et al.
[166] proposed a CV-CNN for land form classification in
InSAR data. Cao et al. [167] addressed CV-FCN for
PolSAR image segmentation. They introduced CV
downsampling - CV upsampling to achieve pixel-wise
classification. Moreover, they proposed a CV form of
cross-entropy loss function and a weight initialization for
their FCN network. Yu et al. [168] proposed CV-CNN for
the SAR-ATR task in which they used only convolutions
instead of pooling and fully connect layers to prevent
overfitting. Wang et al. [169] proposed a CV-CNN for forest
height mapping using polarimetric interferometric
SAR (PolInSAR) data. Li et al. [170] combined CV-CNN
with Contourlet filter bank to extract contour, edge and
texture information for PolSAR image classification.
Moving targets appear defocused and azimuthally displaced
when using conventional SAR image formation
algorithms. Mu et al. [171] addressed the problem ofmoving
targets by using a CV-CNN, namely DeepImaging.
Their proposed CV-CNN-based framework is depicted in
IEEE A&E SYSTEMS MAGAZINE
29
IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV
