IEEE - Aerospace and Electronic Systems - December 2019 - 34

Cognitive Radar for Classiﬁcation

Figure 8.
Basic structure of an AE.

Figure 7.
CNN with SVM classifier applied to ISAR images.

with deeper networks, which posses a larger number of
free parameters and therefore a larger capacity, an overfitting effect occurs and the network starts to remember the
training data and is not focusing on common features of
different classes. The network in Figure 7 is a five-layer
network, but the last two layers have been removed and
replaced by SVMs. In this way, only the feature extraction, which results in the internal representation stored in
the third layer of the CNN, is used for classification. This
automatic feature extraction is known as representation
learning [26] and a big strength of deep learning networks,
since no hand-crafted features are necessary for the
classification.
As an example for the classification performance of this
classifier, some of the results from [23] are shown here. An
ensemble of ten networks of the original five layer structure
with four feature maps in the first layer, twelve elements in
the third layer, i.e., the internal representation, and 24 neurons in the fourth layer for classification achieved a correct
classification rate of 93.23% in average. With the use of
SVMs this already good result could be improved to 94.38%.
For this experiment five different targets with a total of 1085
images were used. These images were separated into 558
training and 527 test images. For four of the five classes
measurements from different flights were available to
achieve a high independence between training and test data.
For the fifth class only data from one flight was available, but
the data was separated into images from different circles that
were flown during the measurement. To further increase the
amount of training data and equalize the distribution of the
different classes, several transformations were used to create
artificial data from the available data. Details about these
34

transformations, further examples and more details about
CNNs in combination with SVMs can be found in [22].
Another feature that is only mentioned here without
example is the use of priority classes, which uses an extra
factor during the training for certain classes to give a reliable classification of these classes. This comes of course
with the cost that the error rate of other classes increases,
but the results of [27] show that the overall classification
rate might increase until an optimum point is reached.

CHANGE DETECTION WITH AES
The algorithm presented in this section was recently proposed [28] and is based on the idea to use an AE to learn a
certain background of a scene and use variations of the
resulting code during the measurement to indicate the
presence of targets. The basic structure of an AE is shown
in Figure 8, which shows a separation into two parts, the
encoder and the decoder. The output of the AE should be
identical with the input in the ideal case and the actual
useful part of the AE is the produced code, which represents the output of the encoder. This code should contain
all relevant information about the input data.
For details about the training method and the internal
structure of the AE, the reader is referred to [28], here
only the results are repeated.
For the training, i.e., the learning of the background,
2042 range profiles are available and additionally to these
real measurements 6000 profiles generated by smoothed
white noise are used for the training of the AE. These
noise profiles are in the dynamic range of the radar and
are necessary to allow proper reconstruction of unknown
profiles that might contain targets with a higher RCS than
the targets in the training profiles. For the actual detection
in the unknown profiles only the difference between the

IEEE A&E SYSTEMS MAGAZINE

DECEMBER 2019

IEEE - Aerospace and Electronic Systems - December 2019

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - December 2019