IEEE - Aerospace and Electronic Systems - February 2023 - 8

Anti-Interference Recognition for Aerial Infrared Object Based on Convolutional Feature Inference Network
Figure 5.
Images of the object and artificial interference regions after preprocessing. (a) Object. (b) Artificial interference.
the simulated dataset including the confrontation images of
fighters and missiles are used in the experiment. Meanwhile,
advanced infrared modeling and simulation technology is
exploited to generate different ballistic data that is close to
the one in the actual air combat confrontation environment.
The relevant experimental parameters are as follows. The
projection distance is 6000 m; the number of interference
release is 12; the interval of the missiles is 0.5 s, and the
entry angle (the angle between the missile body and the axis
of the object aircraft when the missile is launched) includes
the ballistic trajectory from 0 to 360 at an interval of 15.
After the seeker image sequence is extracted, the object and
the interference regions are labeled. Then, these regions are
used as the positive and negative samples to construct the
datasets shown in Figures 3 and 4.
CONVOLUTIONAL FEATURE EXTRACTION
Before the principal component analysis is performed on
the network feature extraction, the size of each connected
region ofthe input image needs to be normalized and scaled
to 64 64 uniformly. Figure 5 shows images of the object
and artificial interference regions after preprocessing.
Then, the 2DPCANet [22] convolutional layer is applied to
extract the features of the corresponding candidate region.
The parameters of the convolutional layer are set as follows.
The size of the convolution kernel is 9 9; the number
of convolution layers is 2, and the number of
convolution kernels is set to the default, i.e., ½8; 6.
After the training set is preprocessed, all candidate
regions constitute a sample set D ¼fDþ;Dg¼ fI1;I2;
.. . ;INþMg, which includes N positive samples ofthe object
andM negative samples of the interference. According to the
process described in section " Convolutional Feature Inference
Network Recognition Model, " the parameters of the
convolution kernel are learned in an unsupervised manner of
2DPCA. After this, the convolution output O of sample D is
8
obtained. Figure 6 shows the convolution kernel learned from
the training set, and Figure 7 illustrates the final output O1.
Through the comparison of the depth feature
images shown in Figures 7(a) and (b), it can be seen
that the feature difference between the object and the
interference is obvious. Also, the feature description
has a strong accuracy and can be used as a sample for
classification.
In the feature encoding stage, the proposed algorithm
adds the maximum pooling operation ofoverlapping windows,
thus increasing the translation invariance and
reducing the dimensionality for the feature. The window
size is set to 5 5, and the overlap rate is 0.5. Then, the
Figure 6.
Convolution kernel group of the feature extraction network. (a)
The first layer convolution kernel. (b) The second layer convolution
kernel.
IEEE A&E SYSTEMS MAGAZINE
FEBRUARY 2023

IEEE - Aerospace and Electronic Systems - February 2023

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - February 2023