IEEE - Aerospace and Electronic Systems - March 2021 - 72

Automatic Target Recognition on Synthetic Aperture Radar Imagery: A Survey

Figure 7.
10-class SOC 1 SAR ATR performance (best seen in color).

testing all major model variants of BMP2 and T72. From
Figure 7, it is obvious that the governing methods are CNN
and SRC-based, with an exception of method 4-3 (i.e.,
Table 4, method 3). Feature-based methods are also available, but these tend to provide lower acc rates. CNN methods tend to be more appealing than their SRC counterparts,
with the performance gain of the former being in the order
of 4%. Specifically, for a SAR patch size of 128 x 128 pixels,
methods 9-18 and 9-16 present the highest ATR performance demonstrating that partially exploiting structures of
state-of-the-art deep networks used in the visual domain,
i.e., the VGG for both these methods, can be highly beneficial. Interestingly, both 9-18 and 9-16 do not involve any
Transfer Learning but employ the original VGG kernel
weights. However, for smaller SAR images (96x96 pixels),
pure CNN provides a relatively lower acc (method 9-15),
while for an 80 x 80 pixel size, high ATR rates require more
complex solutions such as a decision fusion scheme combining CNN and SRC (method 10-3). CNN-based methods
offer a higher ATR performance than SRC because the latter
downsamples the SAR image to create an overcomplete dictionary discarding some of the target's features, and ultimately diminishes the descriptiveness of the SRC method.
Regarding the feature-based techniques, these are highly
dependent on the originating feature extraction and description concept of each method. However, since each feature
description method was originally designed for images in
the visual domain, the performance of the feature-based
methods for SAR ATR depends on the capability of each
technique to handle equally well the SAR and the visual
data domain.

The next most common 10-class SOC configuration
used in the literature is SOC 2. Compared to SOC-1, this
subset employs only a single target model for BMP2 and
T72 rather than all available ones and is therefore considered of relatively lower complexity. Indeed, regardless of
the ATR method and SAR image size used, the average
acc on SOC-1 is 96.5%, while on SOC-2 99.1%.
Figure 8 clearly shows that deep learning-based algorithms attain a minor performance gain over their counterparts. However, in contrast to the SOC-1 evaluation subset,
here the acc difference is only 0.7% between the CNN and
the SRC-based methods. Additionally, Figure 8 shows that
acc is less related to the SAR image size, as various image
sizes are equally spread throughout the acc performance
plot. Hence, the ATR performance is less affected by the
background correlation as top-performing solutions
involve SAR image patches with various background correlation levels. However, as stated earlier, the complexity
of this target subset is lower allowing higher acc rates.
Regarding the 3-class SOC trials, SOC 1 is the most
evaluated configuration (for details see Table 3). This subset is similar to the 10-class SOC 1, in terms that for the
BMP2, and T72 targets all major models are exploited both
during training and testing. Thus, interestingly, the top-3
performing techniques are the same for both these SOC
subsets, and indeed CNN-based algorithms are still performing relatively better than their competitors, with SRC
and ASC to follow closely (see Figure 9). The CNN methods are more efficient as the fine details of the various training and testing models are properly encrypted by the
convolutional layers. On the contrary, SRC-based pipelines

Figure 8.

Figure 9.

Performance of 10-class SOC 2 (best seen in color).

Performance of 3-class SOC 1 (best seen in color).

72

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
