IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 30

A Survey on the Applications of Convolutional Neural Networks for Synthetic Aperture Radar: Recent Advances
Figure 12.
The proposed CV-CNN based framework in [171] for moving
targets imaging in SAR data.
Figure 12. Many defocused complex SAR images were
used to train their CNN model. The loss function is
defined to compare the real output with the ground truth
image. One drawback of the model in [171] is that the
authors neglected the geolocation task. In fact, they
assumed that radial velocity of moving targets is zero and
only refocusing of the moving targets is solved.
DATA AUGMENTATION
Data augmentation mechanism is widely employed to
expand the diversity of training samples when the available
dataset is small. In EO domain, this technique typically
consists of scaling, cropping, padding, rotation,
flipping, translation and so on. However, SAR data augmentation
must be tied to the sensor itself meaning that
rotation and flipping will not work as it does in EO domain
because of the shadowing issues. Ding et al. [42] proposed
a CNN for the SAR-ATR task with the emphasis on data
augmentation techniques such as adding speckle noise,
translation and pose synthesis. Kwak et al. [172] have also
used speckle noise addition to make their CNN robust for
SAR-ATR tasks. They adopted a similar approach to [42]
i.e., using a filter to reduce speckle and then adding random
speckle generated from a truncated exponential distribution.
They employed Improved Lee sigma (ILS) filter [173]
instead of median filtering that has been used in [42]. Du
et al. [174] proposed a displacement- and rotation-insensitive
CNN for data augmentation in MSTAR dataset. Their augmented
images are shown in Figure 13 as an example ofdata
augmentation in SAR images. Lv and Liu [175] used the concept
ofattributed scattering centers (ASC) for data augmentation.
ASCs, which had already been employed for SAR-ATR
tasks in 1990s by Potter and Moses [176], represent the electromagnetic
scattering characteristics oftargets. Note that the
scattering centers are very much affected by the variation of
the depression angle and thus the majority techniques relying
only on ASC have poor robustness in experiments with
extended depression angle variation [177]. Sparse representation
was used in [175] to extract the ASCs of a single SAR
image. Subsequently, some of the ASCs were selected for
reconstruction of the image and many new training samples
were generated by repeating this procedure. In fact, the overcomplete
dictionary in sparserepresentation, smoothens the
local radar reflectivity parameters and ultimately increases
the robustness towards the depression angle variation [177].
Jiang and Zhou [178] have also proposed a hierarchical
fusion ofASC matching and CNN for a more robust performance
in SAR-ATR applications.Wang et al. [179] proposed
a generative model based on Wasserstein AE [180] using
convolutional layers, fully connected layers and residual
modules that generate SAR images. Their model is
useful for augmenting the dataset to increase the recognition
accuracy.
SPARSE CODING AND CNN
Sparse coding (SC) is used for learning how to efficiently
represent the input data by using a linear combination of
Figure 13
Data augmentation technique employed in [174].
30
IEEE A&E SYSTEMS MAGAZINE
MAY 2022

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