IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 23

Oveis et al.
Figure 3.
Proposed SAR-ATR architectures. (a) Combining CNN and SVM [44]. (b) Multiview [59].
SAR-ATR applications. Chen et al. [55] used weight
quantization and Huffman coding to compress a CNN for
resource-constrained SAR-ATR tasks. Min et al. [56] proposed
a micro CNN, which has only two layers with
weights defined as " 1, " " 1, " and " 0, " for decreasing the
inference time of SAR-ATR tasks. Comparing with [55],
the authors of [56] showed that their proposed method
achieves a low error rate with a smaller model size and
fewer parameters to process. The proposed technique in
[56] is based on knowledge distillation [57], also known
as " teacher-student paradigm, " for training the microCNN
under the supervision of a deeper network. Zhang
et al. [58] have also studied knowledge distillation for
real-time SAR-ATR tasks.
Some studies have discussed the applicability ofCNNs
for multiview SAR-ATR, which benefits from diverse recognition
information by acquiring SAR images from different
viewing angles. Pei et al. [59], [60] suggested a
multiple-input CNN architecture, termed " multiview, " to
learn the features of targets at different azimuth angles.
Their proposed architecture together with the one proposed
by [44] are depicted in Figure 3. Similarly, Zhao et al. [61]
proposed a framework named as multistream CNN for
SAR-ATR tasks. Their suggested multi-input architecture
is fed by multiple views of the same target. Afterward, a
Fourier feature fusion layer has been added to fuse the features
derived from multiple views.
In some SAR-ATR studies, feature fusion and aggregation
have been addressed. Wang et al. [62] proposed a
two-channel CNN to extract and fuse features from the
intensity and edge information of SAR images for SARATR
tasks. Cho et al. [63] proposed a framework based
MAY 2022
on an aggregation of three feature extractor subnetworks
for SAR-ATR application. Their work relies on the parallel
use of max-pooling and average-pooling operations.
Max-pooling extracts strong features from images. On the
contrary, average-pooling is not often used because it
makes features smooth. However, it can reduce the effect
of noise in SAR images [63]. Bai et al. [64], proposed a
CNN-based architecture for ATR of multipolarized ISAR
images based on the fusion of the features from HH, HV,
and VV-polarized ISAR images. This was done by adopting
a double-layer spatial transformer network (STN) [65]
module to adjust the image deformation ofeach polarimetric
channel. STNs constitute a widely used end-to-end
trainable solution for CNNs to learn invariance to image
transformations [66].
Insufficient SAR labeled data has been mentioned in
the major part of the papers in SAR-ATR. Apart from data
augmentation methods, which will be covered in " Related
Topics to CNN and SAR " , there are some other alternatives
to obviate this problem. Lin et al. [67] proposed a
convolutional highway network to train deeper networks
with limited SAR data. Highway networks [68], which are
based on CNNs, can be trained with smaller datasets by
exploiting skip connections, similar to Res-Net, for training
deep models. Yue et al. [69] proposed a semi-supervised
CNN method for SAR-ATR in which the
information contained in the unlabeled samples was integrated
into the loss function by means of a linear discriminant
analysis method. In fact, they utilized a CNN to
obtain the class probabilities of the unlabeled samples.
As previously mentioned, MSTAR benchmark dataset
[8] is one of the few publicly available labeled dataset for
IEEE A&E SYSTEMS MAGAZINE
23
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
