IEEE - Aerospace and Electronic Systems - May 2022 - Tutorial XV - 31

Oveis et al.
atoms from an overcomplete dictionary. In other words,
the main goal of sparse coding is to find a sparse representation
from an overcomplete basis set [181]. Some studies
addressed the combination of SC and CNN to solve a
SAR task. SC-ATR searches for linear projections
between the target and the feature spaces, while CNNATR
searches for nonlinear projections [182]. KechagiasStamatis
and Aouf [182] opened a new research area
to improve SAR-ATR results by fusing these two concepts,
i.e., SC and CNN. Instead of data-level and feature-level
fusion schemes, they implemented a
decision-level fusion. The reason why they ignored the
data-level fusion is that both the CNN and the SC
modules use the same input data in their architecture.
Therefore, any fusion scheme in this sense is meaningless.
Moreover, regarding the processing time requirement
in near real-time applications, they have
neglected feature-level fusion. Gao et al. [183] combined
CNN and Joint Sparse Representation (JSR) for
SAR target classification. They fused deep feature vectors
from different convolution layers based on the
multi-canonical correlation analysis (MCCA) [184] to
maintain the relevance and eliminate the redundancy.
Subsequently, labels were provided for targets based
on the JSR error. JSR, which is a natural generalization
of the single-task sparse representation, is a multitask
problem and it considers the inner correlations
between different tasks [185], [186]. It is worthwhile
to mention that Zhang et al. [187] had already
addressed the concept of multi-view SAR-ATR using
JSR in 2012, however, they did not exploit CNN. Lv
et al. [188] have also employed feature maps from
different convolution layers to generate multi-level
deep features. Afterwards, they used JSR for the classification
of generated multi-level deep features. Li et al.
[189] employed sparse coding to reduce the computation
burden and memory occupation of a sliding window
FCN for PolSAR image classification.
TRANSFER LEARNING
Transfer learning is used to solve new tasks of which we
do not have sufficient data for training, by transferring the
knowledge from previous successful models. Transfer
learning focuses on borrowing knowledge from one task
(source domain) to another related task (target domain).
One approach of transfer learning is to reuse a pretrained
model. A pretrained model, as already trained on a large
dataset, can be employed again to solve a new but similar
problem. Transfer learning can decrease the required time
for training procedures and increase the accuracy of the
model when the labeled data is not sufficient for training.
It is typically done by substituting the last layers of the
network by new layers to fit the new problem. Finally,
MAY 2022
based on the size of the new dataset and its similarity to
the older one, the entire model, or only some layers, will
be fine-tuned. A typical case is the use ofCNNs pretrained
by optical images and then applied to SAR images.
However, reusing well-known pretrained networks
does not always achieve satisfactory performance for
SAR applications since there exists a prominent discrepancy
between SAR and optical images [190].
Instead, some studies followed different approaches
such as transferring knowledge from simulated SAR
data, unlabeled SAR images, and so on. Hansen et al.
[191] showed that a CNN, pretrained on simulated
data outperforms the one that is trained only on real
data, especially when the labeled real data is not sufficient.
They transferred knowledge from a simulated
SAR dataset and fine-tuned it by using MSTAR dataset.
Similarly, Wang et al. [192] utilized transfer learning
between simulated SAR data and real SAR data to
solve the problem of insufficient training samples.
They used adversarial domain adaptation [193] to handle
the problem of domains shift between the source
and the target datasets. Huang et al. [194] proposed an
assembled CNN architecture, including a reconstruction
pathway and a classification pathway, for transferring
the knowledge from a large number of unlabeled
SAR images. In other words, they used unlabeled SAR
images to train the reconstruction pathway and then
reused the pretrained convolutional layers for SAR
image-based classification. Their proposed framework
is depicted Figure 14. Cui et al. [195] used unlabeled
SAR images to feed both a CNN and an assistant classifier
(SVM).
Afterwards, those unlabeled samples whose category
recognition confidence is higher than a certain threshold
are fed to the CNN again for fine tuning. Sun et al. [196]
studied transferring knowledge about aspect angles of
ground target from the source domain to the target
domain. They proposed an angular rotation generative network
to tackle the lack of training data at different aspect
angles, which inevitably deteriorates the performance.
Rostami et al. [197] proposed transferring knowledge
from EO domain to the SAR domain using the Sliced
Wasserstein Distance (SWD) [198] to measure and then
minimize the discrepancy between the source - target
domains. Some studies have also benefited from wellknown
CNN architectures that have been originally developed
for natural images in the computer vision field.
Zhong et al. [199] transferred the knowledge from CaffeNet
[200], which is a modified version of Alexnet, to solve
SAR classification tasks. Their suggested transfer-learning
scheme is depicted in Figure 15 where they have
employed MSTAR dataset [8] as well. Wang et al. [201]
transferred a partial structure of VGGNet pretrained on
the ImageNet data for the target detection in SAR images.
Since the colorful natural images have three channels,
IEEE A&E SYSTEMS MAGAZINE
31
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
