IEEE - Aerospace and Electronic Systems - February 2023 - 6

Anti-Interference Recognition for Aerial Infrared Object Based on Convolutional Feature Inference Network
scale, shape, grayscale changes, and anti-interference concealment
is ignored. Similar problems are faced by the antiinterference
recognition ofaerial infrared objects.
The object recognition method based on feature
extraction has become mainstream in the object recognition
field. Two issues are essential to the method. The first
one is finding features with better characterization ability
and realizing the automation of feature extraction. The
other one is improving the robustness of the classification
method. The object recognition method proposed in this
article is based on a convolutional feature inference network.
First, it takes the regions to be classified in the
image as potential objects. Then, it extracts the depth features
of the positive and negative samples from the dataset
and obtains the relationship between each feature. Subsequently,
it constructs the Bayesian network topology
structure and determines the conditional probability table
between each feature node. Finally, based on the Bayesian
classifier theory, the proposed method distinguishes the
object from the interference.
CONVOLUTIONAL FEATURE INFERENCE NETWORK
RECOGNITION MODEL
Deep convolutional features are usually obtained by
abstracting features such as contours and textures through
a trained model. The obtained features still mainly express
the contour information of the object. It should be noted
that the contour and texture information of different types
of targets are also different.
Definethe imagematrixas Z. zi;j denotes the element
in the ith row and the jth column of the image. W
represents the weight matrix. wm;n denotes the weight of
the mth row and the nth column in the image. Then,
the feature map of the corresponding point is represented
as
Q ¼
X
m
X
n
wm;n zimþ1;jnþ1:
(1)
Corresponding to the 3 3 weight matrix, the value of
the required feature point is
qi;j ¼
X2
m¼0
X2
n¼0
wm;nziþm;jþn þ wb
(2)
where Q represents the feature image matrix; qi;j represents
the information of corresponding point in the feature
image, and Wb represents the bias term of the convolution
kernel. Using f to represent the activation function and
applying the eigenvalues to the activation function, we
have
qi;j ¼ f
6
!
wm;nziþm;jþn þ wb
X2
m¼0
X2
n¼0
:
(3)
where fXijY:ðxijyi:Þ is the conditional density ofXi when
Y ¼ yi. fYjXi:ðyijxi:Þ is the conditional probability of Y
when Xi ¼ xi.
For the eigenvector A ¼fX1;X2;X3; ... ;Xng, the
Bayesian equation is
fYX1;
j
¼ Pm
j X2;;Xn ðyi x1;x2; .. . ;xnÞ
fX1;X2;;Xn Yj ðx1;x2; ... ;xn yij ÞfYðyiÞ
j¼1 fX1;X2;;Xn Yj ðx1;x2; ... ;xn yij ÞfYðyjÞ
Since the denominator
IEEE A&E SYSTEMS MAGAZINE
Pm
:
(7)
j¼1 fX1;...;XnjY:ðx1; ... ;
xnjyi:ÞfYðyjÞ of (7) is constant for all classes, the class
FEBRUARY 2023
If the depth of the image before convolution is D, the
depth of the corresponding convolution kernel is also D,
and the above equation can be expanded to
qd;i;j ¼ f
!
wd;m;n zd;iþm;jþn þ wb
DX1
d¼0
X2
m¼0
X2
n¼0
(4)
where qd;i;j represents the eigenvalue corresponding to the
dth layer, the ith row, and the jth column. The corresponding
feature map is Qðd; i; jÞ with a dimension ofD 2.
The object features extracted by the deep neural network
are used as high-dimensional features, which have a
different eigenvector dimension from that required by the
Bayesian network. Hence, it is necessary to convert the
high-dimensional deep features into a low-dimensional
eigenvector.
The dimension of the depth feature map Qðd; i; jÞ is
D 2. Denoting the reduced dimension function as ', the
eigenvector X after dimension reduction [19], [20], [21]
can be expressed as
X ¼ 'ðQðd; i; jÞÞ
(5)
where X ¼fX1;X2; ... ;Xng is an n-dimensional vector
that represents the feature descriptor ofone type ofsample
for the depth feature in the Bayesian network.
As a probability model, the Bayesian model has a stable
classification efficiency. By combing prior probability
and subjective probability, the model performs well for
handling multiclassification problems.
Them types ofsamples that need to be classified through
the above method can be obtained as follows. For each type
of sample, n features are selected to form the eigenvector
X ¼fX1;X2 ;Xng,where X1;X2; .. . ;Xn are called
the attribute variables ofthe instances, and they are also continuous
random variables. fXi
ofXi. The class variable Y indicatesmdiscrete random variables
including y1, y2, ,and ym. fYðyiÞ¼ PðY ¼ yiÞ.
For a certain featureXi, the Bayesian equation is
fYXij ðyi xij Þ¼ Pm
fXi Yj ðxi yij ÞfYðyiÞ
j¼1 fXi Yj ðxi yj
ÞfYðyjÞ
(6)
ðxiÞ is the probability density
IEEE - Aerospace and Electronic Systems - February 2023

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