IEEE - Aerospace and Electronic Systems - March 2021 - 68

Automatic Target Recognition on Synthetic Aperture Radar Imagery: A Survey
feature extraction module that considers area, length, slope,
and gray value calculation. The episodic features are further
utilized to enhance the ATR performance by estimating the
aspect angle of the observed target. Training instances with
aspect angles that exceed a Æ10 from the target aspect
angle are discarded during the classification process.

Figure 6.
Typical SAE algorithm. Several AE are stacked layer by layer to
extract deep features. (h denotes the hidden features per layer).

Stacked Autoencoders (SAE). Stacked autoencoders (SAE)
are sequentially placed autoencoders (AE) (see Figure 6),
where the latter is a neural network that comprises an
encoder and a decoder. The encoder remaps the input data
to a hidden representation h according to
h ¼ fðxÞ ¼ sf ðWx þ bh Þ

(12)

where sf is a nonlinear activation function, e.g., sigmoid
or ReLU, W is a weight matrix and bh a bias vector.
Accordingly, the decoder remaps h to a reconstructed version of the input data by minimizing a cost function. In
terms of SAR ATR, only the encoding part on an AE is
used, which is then input to a classification scheme, e.g.,
SVM. AEs are employed in a stacked fashion to increase
descriptiveness.
Kang et al. [14] suggested a feature fusion process that
involves 23 texture-based properties of the SAR image
with features extracted via a three-patch local binary pattern (TPLBP) process. The former features include properties such as area, the bounding box, convex hull, etc.,
while the latter involves a set of classic local binary patterns (LBP) distributed uniformly on a ring of radius r and
one LBP is placed at the center of the ring. Features are
then cascaded and input to an SAE. Deng et al. [13]
reduced overfitting by modifying the reconstruction error
of SAE by adding a Euclidean distance restriction for the
hidden layer features followed by a Dropout layer. Other
autoencoder based solutions are influenced by the human
visual cortical system [123] or are combined with a Synergetic neural network concept [124].

Restricted Boltzmann Machine (RBM). Huang et al. [125] are
inspired by the human cognition system and suggest a convolutional restricted Boltzmann machine (CRBM) algorithm that learns the episodic features which are then used
to extract the semantic features of a SAR image. For the
episodic feature extraction, a convolutional deep belief network is trained to segment the target's SAR reflection and
its shadow. These regions are then input to the semantic
68

Hybrid Deep Learning Techniques. Huang et al. [126] suggested an ATR algorithm that combines the CNN and
SAE concepts along with the transfer learning technique.
The CNN pipeline consists of a classification and a reconstruction module that is used as feedback to estimate the
reconstruction loss. To avoid overfitting, the authors apply
data augmentation and train the reconstruction module of
the CNN pipeline with stacked convolutional autoencoders (SCAE). Chen and Wang [17] also combined the
concepts of SAE and CNN by involving a single layerCNN that is trained on randomly sampled image patches
utilizing an unsupervised sparse autoencoder.
Zhang et al. [127] combined a residual (ResNet) network and a bidirectional long short-term memory
(BiLSTM) network. Fusing these two types of networks
aims at extracting the target's scattering features from a
single-aspect image (ResNet module) and extend their
encoding over several adjacent multiaspect images through
the recurrent learning (BiLSTM module). The latter network is a variant of the popular long short-term memory
(LSTM) layer. For an analysis of the operating principles
of LSTM layers, the reader is referred to [128].

HYBRID REFLECTIVITY ATTRIBUTE-BASED METHODS
Table 10 presents the hybrid reflectivity attribute-based
algorithms. For example, Zhang et al. [129] combined the
concepts of feature-based SAR ATR with deep learning.
Specifically, the authors propose a multiaspect algorithm
where Gabor filters and three-patch local binary patterns
extract comprehensive spatial features, which are then
input to a multilayer perceptron network followed by a
bidirectional LSTM recurrent neural network. Lv and Liu
[130] proposed a novel data augmentation algorithm,
where the ASC are extracted using a sparse representation
(SR) scheme [52], [131]. ASCs are created with various ui
parameters (2), the quality of which is evaluated based on
a quality factor. Finally, the augmented dataset is used to
train a CNN.
Spurred by the classification enhancement of decision
fusion, current literature suggests various strategies of that
type. For example, Karine et al. [132] combined the concepts of local features, saliency maps, and SRC. Specifically, given a SAR image, they create a saliency map
inspired by the human visual system to establish a region
of interest (ROI) type of area containing the target. Then,
the scale-invariant feature transform (SIFT) [133] (a 2-D
local keypoint detector and descriptor) is applied on the

IEEE A&E SYSTEMS MAGAZINE

MARCH 2021

IEEE - Aerospace and Electronic Systems - March 2021

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - March 2021