remote sensing

Article
Structure Tensor-Based Infrared Small Target Detection Method
for a Double Linear Array Detector
Jinyan Gao 1, * , Luyuan Wang 1 , Jiyang Yu 1 and Zhongshi Pan 2

1 Institute of Spacecraft System Engineering, China Academy of Space Technology, Beijing 100094, China
2 Institute of Remote Sensing Satellite, China Academy of Space Technology, Beijing 100094, China
* Correspondence: [email protected]

Abstract: The paper focuses on the mathematical modeling of a new double linear array detector.
The special feature of the detector is that image pairs can be generated at short intervals in one scan.
After registration and removal of dynamic cloud edges in each image, the image differentiation-
based change detection method in the temporal domain is proposed to combine with the structure
tensor edge suppression method in the spatial domain. Finally, experiments are conducted, and
our results are compared with theoretic analyses. It is found that a high signal-to-clutter ratio
(SCR) of camera input is required to obtain an acceptable detection rate and false alarm rate in real
scenes. Experimental results also show that the proposed cloud edge removal solution can be used to
successfully detect targets with a very low false alarm rate and an acceptable detection rate.

Keywords: small moving target detection; double linear array detector; cloud edge removal; structure
tensor

1. Introduction
Citation: Gao, J.; Wang, L.; Yu, J.; Pan, Dim small target detection is a major problem in numerous fields, such as infrared
Z. Structure Tensor-Based Infrared search and track (IRST) systems and external intrusion warnings [1–3]. Since the imag-
Small Target Detection Method for a
ing distance is long in these applications, the target usually occupies only one or a few
Double Linear Array Detector.
pixels [4–6], and there is insufficient texture and shape information for target detection [7–9].
Remote Sens. 2022, 14, 4785. https://
Furthermore, the intensity value of the infrared target is very low due to reflection, refrac-
doi.org/10.3390/rs14194785
tion, the sensor’s aperture diffraction effects and geometric aberrations [10–13]. Therefore,
Academic Editors: Jian Yao, Li Li and it is difficult to separate infrared small targets from complex backgrounds.
Wei Zhang
1.1. Related Works
Received: 4 August 2022
Accepted: 21 September 2022
Existing infrared target detection approaches can be divided into spatial, temporal,
Published: 25 September 2022
and spatio-temporal detection methods. Most spatial detection methods use spatial fil-
tering techniques, and they are usually based on the assumption that the target has a
Publisher’s Note: MDPI stays neutral
larger intensity value than the background. However, this assumption does not always
with regard to jurisdictional claims in
hold in real scenes [14–17]. The temporal detection methods usually use the temporal
published maps and institutional affil-
profiles of each pixel in a sequence of infrared images to extract the small target of interest.
iations.
They have a good detection performance when a small target appears in slowly evolving
backgrounds [18–21]. However, these methods often consume more time than single-frame
detection methods. The spatio-temporal detection methods are complementary to the
Copyright: © 2022 by the authors.
singular spatial or temporal detection methods [22–24]. They use features in both spatial
Licensee MDPI, Basel, Switzerland. (e.g., the gray difference feature) and temporal (e.g., the motion difference feature) domains
This article is an open access article to completely separate targets from clutter.
distributed under the terms and Several optical systems have been proposed to detect small infrared small targets
conditions of the Creative Commons over the past few decades [25–28]. They can be divided into two classes: scanning camera-
Attribution (CC BY) license (https:// based optical systems and staring camera-based optical systems. Scanning cameras have
creativecommons.org/licenses/by/ a relatively wide field of view and are suitable for early warning for large areas [29].
4.0/). However, imaging in this way has a high time delay integration (TDI) in adjacent frames.

Remote Sens. 2022, 14, 4785. https://doi.org/10.3390/rs14194785 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2022, 14, 4785 2 of 13

Consequently, it is difficult to perform data association, and the time for target discovery is
long. On the other hand, the staring camera is usually used for target tracking as its imaging
size is usually small [30], which makes it not suitable for searching in early warning.

1.2. Contributions
In order to overcome these limitations of traditional optical systems, we propose to
use a double linear array detector to detect targets with cross-pixel moving. For a double
linear array detector, two images (an image pair) are generated when the detector scans
from the top to the bottom only once. Figure 1a shows an IR image pair acquired by a
double linear array detector with a slowly changing cloud clutter. It not only reserves the
wide field of view of traditional scanning systems but also reserves the short-time intervals
in adjacent frames of staring systems. A double linear array detector has the following
three advantages. First, it reserves the wide field of view of traditional scanning systems.
Second, its exposure time in each pixel is longer than traditional scanning systems. Third,
the interval between adjacent frames is shorter than traditional scanning systems, which
makes it easy to perform data association subsequently estimate the velocities and relative
positions of cross-pixel moving targets.
Our other task is to automatically detect targets with cross-pixel moving in image pairs
acquired by a double linear array detector. Considering the special spatial arrangement
and imaging modes of the detector (as analyzed in Section 2), the image pair is almost
observed from the same solar angle and atmospheric conditions, and the slow change
in cloud background can be almost negligible at a short interval. Therefore, the image
differentiation-based change detection method is suitable for the detection of targets after
image registration. Figure 1b shows the detection results after image differentiation. We
can see the positive gray-scale value and the negative gray-scale value of a candidate target
produced from an image pair in the filtered result, namely, positive and negative target
pairs in our paper. The mathematical model of image differentiation is:

Dxij = xij (t2 ) − xij (t1 ) (1)

where x (t1 ) represents the image acquired by the first linear array and x (t2 ) represents the
image acquired by the second linear array. i and j indicates the location of the pixel. Dxij
represents the residuals after image differentiation. If the grayscale value of the candidate
target in x (t1 ) is positive, the grayscale value of the candidate target in x (t2 ) is negative
after image differentiation. We call them positive and negative target pairs in this paper.

Target
Target
pos-target

neg-target

(a) (b)

Figure 1. An example of image pair and positive and negative target pairs. (a) An image pair with
slowly changing cloud clutter. (b) Local detection results after image differentiation.

Since false alarms after image differentiation are mainly caused by cloud edges [31,32],
edge suppression has been found to be a useful spatial method complementary to our
temporal change detection method [33–35]. The structure tensor has been widely used
Remote Sens. 2022, 14, 4785 3 of 13

for cloud edge suppression in recent years. Dai et al. [23,24,36,37] allocated the structure
tensor as an adaptive weighting parameter to suppress strong cloud edges. Liu et al. [38]
introduced the gradient direction diversity (GDD) method to suppress sharp cloud edges.
The GDD measure is also inspired by the structure tensor. Li et al. [39] used the local
steering kernel to encode the infrared image patch, as it can represent different intrinsic
structures in different image regions (e.g., the cloud edge region, the flat region, the textural
clutter region and the small target region). Thus, the structure tensor is used for cloud edge
removal in our paper.
The overview of our proposed method is as follows.
• The structure tensor is used to detect infrared small targets, which is used as an
adaptive weighting parameter to suppress strong cloud edges.
• Considering that using information of image sequences requires more prior informa-
tion and a large amount of data processing, the temporal image differentiation filter is
used to extract target pairs using movement information of the target.
• Adaptive thresholding-based constant false alarm (H-CFAR) is performed to obtain
candidate targets, and data association is performed to extract positive and negative
target pairs.
The rest of this paper is organized as follows. Section 2 analyzes the optical path
and mathematical model of the double linear array. Section 3 presents the target detection
model in detail. Section 4 tests the performance of our proposed method. The paper is
concluded in Section 5.

2. The Double Linear Array Detector

For a double linear array detector, the region of interest in the instantaneous field of
view (IFOV) can be measured twice in the same period of time once the detector scans
from the top to the bottom. The optical path of a double linear array detector is shown in
Figure 2. The device consists of a scanning mechanism, an optical system and a focal plane.
The scanning mechanism mainly includes a pendulum mirror and a driving shaft, as wel
as the focal plane constructed with two linear arrays arranged in parallel, as shown in the
right part of Figure 2. Incident light containing the radiation energy information of targets
and the background is first projected to the pendulum mirror; the driving shaft is then
used to rotate the pendulum mirror. After several reflections and refractions in the optical
system, the light finally converges to the focal plane to generate an image pair of the scene.

1. Scanning mechanism 2. Optical system 3. Focal plane

d
……

d
l

Figure 2. The optical path of a double linear array detector.

For a scanning system, the ground sample distance (GSD) determines the maximum
spatial resolution of a camera and the minimum detectable velocity of a target. In addition,
GSD mainly depends on the instantaneous field of view (IFOV) in practice. The IFOV is
the angular cone of visibility of the camera and determines the area on the Earth’s surface
 Remote Sens. 2022, 14, 4785 4 of 13

空军工程大学硕士研究生学位论文
that can be seen from a given altitude at a particular moment [40]. The geometry of the
detecting system, including GSD, IFOV, camera height (H), pixel size (d), and the focus
of the optical system ( f ) is shown in Figure 3, and the relationship between them can be
首先，分析弹道目标在图像中的尺寸大小。通常弹道导弹尾焰的宽度在十几
expressed as:
IFOV Hd
GSD = 2H tan( )= (2)
米左右，长度能达到一百米以上，在真空中洲际弹道导弹的尾焰甚至长达上千米。 2 f
Considering that our detection method is based on two frames, and it has to use the
要想分析弹道目标在图像中的尺寸大小，得先分析探测系统的空间分辨率，即探
distance constraint to associate the positive and negative target pairs, the following data
association constraints are derived.
测系统能分辨物体的空间几何长度的最小极限，在图像中表示为一个像元所代表
The target velocity in the image plane v pixel is used to predict the target velocity
v in real scenes. Since the dual linear array can only detect cross-pixel moving targets,
the detectable velocity in the focal plane is:
的地面范围的大小。
d l
<v < (3)
卫星的空间分辨率示意图如图 2.3 所示。在某一瞬间，探测器的探测单元所对
∆t pixel
∆t
where d is the pixel size, l is the width of the linear array, and ∆t is the time interval of
应的视场角称为瞬时视场角（Instantaneous
linear arrays. Field Of View, IFOV）。其在地面上对应
Then, the range of the detectable target velocity in real scenes is:
的尺寸大小称为地面分辨率单元（Ground Resolution Cell, GRC），即卫星的空间分
GSD l GSD
<v< · (4)
∆t ∆t d
辨率，它可表示为：
The left and right sides of Equation (4) are the minimum and maximum detectable
velocity of the target in real
IFOV Hd are denoted by vmin and vmax , respectively.
scenes, which
 2 H tan(
GRCEquations
From  can obtain the distance constraints
(3) and (4), )we （2.1）
as follows:
2 f
vmin · ∆t vmax · ∆t
< ∆D < (5)
GSD GSD
式中， H 为探测系统到目标的作用距离。 d 为像元尺寸， f 为光学系统的焦距。
where ∆D is the target moving distance in linear arrays.
d

f
focal point

IFOV
H

Horizon line
GSD
the earth
Figure 3. Geometry of the detecting system.

3. Target图2.3 空间分辨率示意图
Detection Model
The proposed small target detection method for a double linear array detector is
shown in Figure 4. We first use the structure tensor method to suppress cloud edges.
由于弹道导弹在助推段内的飞行高度较低，远小于卫星距离地面的高度，可
Then, the temporal image differentiation filter is used to extract target pairs using motion
information of the target. After background suppression, adaptive thresholding-based
以忽略导弹的飞行高度。卫星的空间分辨率与卫星距地面的高度以及卫星的视角
constant false alarm (H-CFAR) is performed to obtain candidate targets. Finally, data

有关，卫星离地面越远，分辨率越低，卫星视角越倾斜，瞬时视场角不变，但观

测面积增大，分辨率降低。
[74]
Remote Sens. 2022, 14, 4785 5 of 13

association is performed to extract positive and negative target pairs using the constraints
given in Equation (5).

Cloud edge removal

Target
Target

Image pairs image differentiation Adaptive thresholding Data association

Figure 4. Overall flowchart for a double linear array detector.

The structure tensor is proposed based on the edge shock filter and variational func-
tionals [41–43]. It turns out to be very effective for the enhancement of corner structures
and presents different characteristics on homogeneous regions, edges, and texture regions
of an image [44,45]. Therefore, it is used in our infrared small target detection method for
cloud edge suppression in a single image.
The structure tensor is essentially a steering matrix [46,47]. It describes the local
structural information about the image, and it can be represented as:

∂I 2
" #
∂I ∂I
Ci = ∑ ∂xi1
∂I ∂I
∂x1 ∂xi2
∂I 2
(6)
xi ∈ Ωi ∂xi1 ∂xi2 ∂xi2

where I represents the image and xi = ( xi1 , xi2 ) represents the two-dimensional coordinate
vector of the central pixel in a rectangular window Ωi .
The steering matrix captures the principal directions of local texture from the gradient
distribution in a small neighborhood (mostly 5 × 5 [48]). Therefore, the structure tensor Ci
can be first calculated by GiT Gi with:
 
Zx1 ( x 1 ) Zx2 ( x1 )
Gi = 
 .. .. 
(7)
. . 
Zx1 ( x P ) Zx2 ( x P )

where Zx1 (·) and Zx2 (·) denote the first derivatives along the horizontal and vertical axes,
respectively, and P is the number of pixels in the local window Ωi . However, since it is
difficult to calculate the gradient distribution, the covariance matrix Ci can be estimated by
singular value decomposition [22,49] as:

Ci = γi Uθi Λi UθTi (8)

where γi is the scaling parameter. It is large in homogeneous regions but small in textured
regions. θi is the rotation parameter; it defines the dominant orientation angle, Uθi , as a
rotation matrix. Λi is the elongation matrix.

00
!1
2
s1 s2 + λ
γi = (9)
M
Remote Sens. 2022, 14, 4785 6 of 13

 
cos θi sin θi
Uθi = (10)
− sin θi cos θi
 
σi 0
Λi = (11)
0 σi−1
0
s +λ
σi = 1 0 (12)
s2 + λ
0 00
where σi is the elongation parameter. λ = 1; λ = 10−1 . The structure tensor can be
calculated using Equations (9)–(12).
The eigenvalues of singular value decomposition is denoted as λ1 and λ2 . They can
be used as two features to describe the local structural information [50]. The larger λ1 is
then λ2 ; the measurement region is more likely a cloud edge region. Therefore, the cloud
edge suppression measure can be defined as follows:

K = exp(−h · (λ1 − λ2 )) (13)

Figure 5 shows the cloud edge suppression results on images with three different
shapes of clouds. It demonstrates that the proposed structure tensor-based measure
achieves good performance in cloud edge suppression.

(a) (b) (c)

Figure 5. Illustrations of cloud edge suppression based on structure tensor measures. (a) Ragged
cloud edge; (b) strong cloud edge; (c) fluffy cloud edge.

Temporal Differentiation of Image Pairs

Assuming that the noise in image pairs follows a Gaussian distribution N (0, σ2 ) and
the registration error is zero, then the residual after image differentiation follows a Gaussian
distribution N (0, 2σ2 ). Since the grayscale value of the residual image can be positive or
negative, we use a bilateral filter to extract candidate points, the positive threshold th_p
is used to extract positive candidate points, and the negative threshold th_n is used to
extract negative candidate points. If both the positive and negative grayscale values of the
target exceed the defined thresholds, the data association step is followed (according to
Equation (5)). Finally, the detection rate pd and false detection rate p f can be calculated by
Equations (14) and (15).

−
pd = p+


 d · pd 
th_p−( Tmax −µ BG )

p+
d = Q √ (14)
 2σBG 
th_n−(µ BG − Tmax )

−
pd = 1 − Q

 √
2σBG

−
where p+ d and pd are the target detection rate with a positive gray value and the target
detection rate with a negative gray value, respectively. Tmax is the highest intensity value in
the target region, and µ BG and σBG are the average and the standard deviation of intensity
Remote Sens. 2022, 14, 4785 7 of 13

values in a residual image. Q( x ) is a right-tailed distribution function, as defined in

Equation (16). 
 p f = N · p+ · p−f
 f

 
√th_p

p+f = Q (15)
 2σ
BG 
 −
 p = 1−Q √
 th_n
f 2σBG

where p+
f and p−
f are the target false alarm rate with a positive gray value and the target
false alarm rate with a negative gray value, respectively. N is the number of pixels that are
possibly associated.
( )
(ξ − µ BG )2
Z +∞
1
Q( x ) = q exp − 2
dξ (16)
x 2πσ2 2σBG
BG

To ensure that both positive and negative targets can be detected, the signal-to-clutter
ratio (SCR) is defined as:
| Tmax − µ BG |
SCR = (17)
σBG
Combining Equations (14)–(17), the relationship between the detection rate, false rate
and SCR can be expressed as:
 1  1 SCR
−1 2
Q p f /N − Q−1 ( pd2 ) = √ (18)
2

The theoretic receiver operating characteristic (ROC) curves validated through Monto
Carlo simulations for the double linear array detector are analyzed as shown in Figure 6. It
can be seen that the detection probability becomes higher as the SNR increases. Specifically,
When the SNR is 6.1 and the false alarm rate is 1 × 10−4 , our double linear array detector
can achieve a detection rate of 97%; when the SNR is above 6.1 and the false alarm rate is
1 × 10−5 , our double linear array detector can achieve a detection rate of 93%.
pd

SNR=6
SNR=6.1
SNR=7
SNR=8

pf

Figure 6. The theoretic ROC curves for the double linear array detector.

4. Experimental Results and Discussions

4.1. Simulation Scenes
In this section, experiments are conducted to test the performance of our method.
Given an image, another image was generated with one pixel jittering in an arbitrary
direction to simulate the image pair produced by the double linear array detector. Then,
simulated targets were added into each image, the target position and motion direction
Remote Sens. 2022, 14, 4785 8 of 13

were randomly determined, the intensity of the targets was determined according to a
specific SCR, and the target velocity was set to 2∼3 km/s. Next, the camera resolution was
set to 1 km × 1 km, and the interval of an image pair was set to 2 s. Finally, considering that
the target can move along the diagonal direction and the horizontal direction, the distance
constraint for a target pair after clutter suppression is set to 2∼7 pixels, as shown in Figure 7.

(a) (b) (c)

Figure 7. Schematic diagram of distance constraint for a target pair. (a) Horizontal direction, (b) Ver-
tical direction, (c) Diagonal direction.

4.2. Experiments on Simulation Scenes

Three experiments were conducted on scenes with different shapes of cloud (i.e.,
Figure 8a–d, ragged cloud scenes; Figure 8e–h, strong cloud scenes; Figure 8i–l, fluffy cloud
scenes), as shown in the first column of Figure 8. The second column in Figure 8 shows
the filtered results after image differentiation, and the third column shows the filtered
results after cloud edge suppression and image differentiation. It can be seen from the
third column of Figure 8 that false alarms in dynamic cloud edges have been significantly
reduced. Finally, accurate association results are obtained, as shown in the fourth column
of Figure 8.

4.3. ROC Curves Evaluation

The ROC curves obtained through Monto Carlo simulations were used to test our
proposed spatio-temporal method for a double linear array detector. An ROC curve
represents the probability of detection Pd as a function of the false alarm rate Pf . Pd and Pf
can be calculated as:
nt
Pd = (19)
nc
nf
Pf = (20)
n
where nt , nc , n f and n represent the number of detected true pixels, ground-truth target
pixels, false alarm pixels and the total number of image pixels, respectively.
As shown in Figure 9, at the cost of an increase in the false alarm rate, the detection
probability is increased under a certain SCR. When SCR is increased, the detection prob-
ability is increased under a certain false alarm rate. It can be observed that the detection
rate and the false alarm rate are a pair of contradictory variables. Furthermore, the theoreti-
cal ROC results are better than the results achieved by other two methods in real scenes.
That is because the residual after background suppression does not completely follow a
Gaussian distribution.
Remote Sens. 2022, 14, 4785 9 of 13

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

Figure 8. An example of results achieved by the proposed small target detection methods on three
images. The red and blue boxes in each image represent the candidate points passing the positive
and negative thresholds (as mentioned in Section 3), respectively. (a–d) Ragged clouds; (e–h) strong
clouds; (i–l) fluffy clouds.

Table 1 shows the probability of detection in three experiments under different SCRs
and Pf . Table 2 shows the false alarm rates in three experiments under different SCRs
and Pd . We can see that with the cloud edge suppression method, the detection rate is
significantly improved, and the false alarm rate is largely reduced.

Table 1. Pd achieved by three methods under different SCRs and Pf .

Pf = 5 × 10−6 Pf = 10−5
SCR Temporal Spatial-Temporal Theory Temporal Spatial-Temporal Theory
6 11.39% 54.40% 88.98% 44.03% 63.93% 91.90%
7 65.53% 74.16% 99.02% 70.67% 81.06% 99.38%
8 84.93% 86.53% 99.97% 88.39% 92.74% 99.98%
‘Temporal’ stands for the temporal image differentiation method; ‘Spatial-temporal’ stands for the method
combining the spatial cloud edge suppression and the temporal image differentiation method.

Table 2. Pf achieved by three methods under different SCRs and Pd .

Pd = 85% Pd = 90%
SCR Temporal Spatial-Temporal Theory Temporal Spatial-Temporal Theory
6 3.30 × 10−4 2.32 × 10−4 3.00 × 10−5 8.82 × 10−4 5.14 × 10−4 6.00 × 10−6
7 4.34 × 10−5 2.10 × 10−5 2.00 × 10−7 6.88 × 10−4 6.16 × 10−5 4.00 × 10−8
8 4.97 × 10−6 4.28 × 10−6 8.00 × 10−10 1.40 × 10−5 8.08 × 10−6 5.00 × 10−9
‘Temporal’ stands for the temporal image differentiation method; ‘Spatial-temporal’ stands for the method
combining the spatial cloud edge suppression and temporal image differentiation method.
Remote Sens. 2022, 14, 4785 10 of 13

ROC ROC
1 1

0.9 0.9

0.8 0.8

Probability of detection (Pd)

Probability of detection (Pd)

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
SNR=6 SNR=6
0.1 SNR=7 0.1
SNR=7
SNR=8 SNR=8
0 −5 −4 −3 0 −5 −4 −3
10 10 10 10 10 10
False alarm rate (Pf) False alarm rate (Pf)

(a) (b)

ROC
1

0.9

0.8
Probability of detection (Pd)

0.7

0.6

0.5

0.4

0.3

0.2
SNR=6
0.1 SNR=7
SNR=8
0 −5 −4 −3
10 10 10
False alarm rate (Pf)

(c)

Figure 9. ROC curves achieved by three methods under different SCRs. (a) Performance achieved
by temporal image differentiation. (b) Performance achieved by spatial cloud edge suppression and
temporal image differentiation. (c) Performance achieved by theoretic analysis.

4.4. Evaluation of the Proposed Detection Framework

The data association results for positive and negative targets under different operating
conditions are shown in Figure 10. From Figure 10a–p, we can see that after image differ-
entiation and cloud edge suppression, the target pairs can be correctly associated, and no
false detection is found in the association results. From Figure 10q–t, we find that one target
can associate with two or more targets under a few conditions. To address this problem,
we have to extend the dual linear array to a multi-linear array in future, and the false
association can be removed by the direction of the target motion trajectory. Furthermore,
the input SCR of the camera can be improved to handle this problem.
Remote Sens. 2022, 14, 4785 11 of 13

(a) (b) (c) (d)

(e) (f) (g) (h)

(i) (j) (k) (l)

(m) (n) (o) (p)

(q) (r) (s) (t)

Figure 10. Thedata association results with targets in different positions in various complex back-
grounds. (a–p) target pairs were associated correctly, (q–t) one target was associated with two or
more targets.

5. Conclusions
This paper has presented a complete framework for small infrared target detection
using a double linear array detector. First, a new double linear array detector was modeled
to generate image pairs at short intervals. Second, considering the limitations of singular
spatial or temporal detection methods, an image differentiation-based change detection
method in the temporal domain was proposed combined with the structure tensor edge
suppression method in the spatial domain. The experimental results showed that targets
can be extracted accurately with a very low false alarm rate and an acceptable detection rate.
Remote Sens. 2022, 14, 4785 12 of 13

Author Contributions: Conceptualization, J.G. and Z.P.; methodology, J.G.; software, J.G.; validation,
J.G. and J.Y.; formal analysis, J.G.; investigation, J.G.; resources, L.W.; data curation, Z.P.; writing—
original draft preparation, J.G.; writing—review and editing, J.G.; visualization, J.G.; supervision,
J.G.; project administration, L.W. All authors have read and agreed to the published version of
the manuscript.
Funding: This research received no external funding.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Zhang, M.; Dong, L.; Ma, D.; Xu, W. Infrared target detection in marine images with heavy waves via local patch similarity.
Infrared Phys. Technol. 2022, 125, 104283. [CrossRef]
2. Rawat, S.S.; Verma, S.K.; Kumar, Y. Infrared small target detection based on non-convex triple tensor factorisation. IET Image
Process. 2021, 15, 556–570. [CrossRef]
3. Zhu, Q.; Zhu, S.; Liu, G.; Peng, Z. Infrared Small Target Detection Using Local Feature-Based Density Peaks Searching. IEEE
Geosci. Remote Sens. Lett. 2022, 19, 6507805. [CrossRef]
4. He, X.; Ling, Q.; Zhang, Y.; Lin, Z.; Zhou, S. Detecting Dim Small Target in Infrared Images via Sub-Pixel Sampling Cuneate
Network. IEEE Geosci. Remote Sens. Lett. 2022, 19, 3189225. [CrossRef]
5. Lu, R.; Yang, X.; Li, W.; Fan, J.; Li, D.; Jing, X. Robust infrared small target detection via multidirectional derivative-based
weighted contrast measure. IEEE Geosci. Remote Sens. Lett. 2020, 19, 7000105. [CrossRef]
6. Yang, X.; Wu, T.; Wang, N.; Huang, Y.; Song, B.; Gao, X. HCNN-PSI: A hybrid CNN with partial semantic information for space
target recognition. Pattern Recognit. 2020, 108, 107531. [CrossRef]
7. Bai, X.h.; Xu, S.w.; Guo, Z.x.; Shui, P.l. Floating Small Target Detection Based on the Dual-polarization Cross-time-frequency
Distribution in Sea Clutter. Digital Signal Process. 2022, 129, 103625. [CrossRef]
8. Liu, T.; Yang, J.; Li, B.; Xiao, C.; Sun, Y.; Wang, Y.; An, W. Nonconvex Tensor Low-Rank Approximation for Infrared Small Target
Detection. IEEE Trans. Geosci. Remote Sens. 2021, 60, 5614718. [CrossRef]
9. Gao, J.; Guo, Y.; Lin, Z.; An, W.; Li, J. Robust infrared small target detection using multiscale gray and variance difference
measures. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 5039–5052. [CrossRef]
10. Huang, Z.; Zhang, Y.; Yue, X.; Li, X.; Fang, H.; Hong, H.; Zhang, T. Joint horizontal-vertical enhancement and tracking scheme for
robust contact-point detection from pantograph-catenary infrared images. Infrared Phys. Technol. 2020, 105, 103156. [CrossRef]
11. Gao, Z.; Dai, J.; Xie, C. Dim and small target detection based on feature mapping neural networks. J. Vis. Commun. Image
Represent. 2019, 62, 206–216. [CrossRef]
12. Ding, L.; Xu, X.; Cao, Y.; Zhai, G.; Yang, F.; Qian, L. Detection and tracking of infrared small target by jointly using SSD and
pipeline filter. Digit. Signal Process. 2021, 110, 102949. [CrossRef]
13. Wang, N.; Li, B.; Wei, X.; Wang, Y.; Yan, H. Ship detection in spaceborne infrared image based on lightweight CNN and
multisource feature cascade decision. IEEE Trans. Geosci. Remote Sens. 2020, 59, 4324–4339. [CrossRef]
14. Li, Y.; Li, Z.; Shen, Y.; Guo, Z. Infrared Small Target Detection Via Center-surround Gray Difference Measure with Local Image
Block Analysis. IEEE Trans. Aerosp. Electron. Syst. 2022. [CrossRef]
15. Zuo, Z.; Tong, X.; Wei, J.; Su, S.; Wu, P.; Guo, R.; Sun, B. AFFPN: Attention Fusion Feature Pyramid Network for Small Infrared
Target Detection. Remote Sens. 2022, 14, 3412. [CrossRef]
16. Zhang, T.; Peng, Z.; Wu, H.; He, Y.; Li, C.; Yang, C. Infrared small target detection via self-regularized weighted sparse model.
Neurocomputing 2021, 420, 124–148. [CrossRef]
17. Wang, H.; Li, H.; Zhou, H.; Chen, X. Low-altitude infrared small target detection based on fully convolutional regression network
and graph matching. Infrared Phys. Technol. 2021, 115, 103738. [CrossRef]
18. Chen, G.; Wang, W.; Tan, S. IRSTFormer: A Hierarchical Vision Transformer for Infrared Small Target Detection. Remote Sens.
2022, 14, 3258. [CrossRef]
19. Qin, Y.; Bruzzone, L.; Gao, C.; Li, B. Infrared small target detection based on facet kernel and random walker. IEEE Trans. Geosci.
Remote Sens. 2019, 57, 7104–7118. [CrossRef]
20. Zhao, M.; Li, L.; Li, W.; Tao, R.; Li, L.; Zhang, W. Infrared small-target detection based on multiple morphological profiles. IEEE
Trans. Geosci. Remote Sens. 2020, 59, 6077–6091. [CrossRef]
21. Yang, P.; Dong, L.; Xu, W. Detecting Small Infrared Maritime Targets Overwhelmed in Heavy Waves by Weighted Multidirectional
Gradient Measure. IEEE Geosci. Remote Sens. Lett. 2021, 19, 3080389. [CrossRef]
22. Zhang, T.; Wu, H.; Liu, Y.; Peng, L.; Yang, C.; Peng, Z. Infrared small target detection based on non-convex optimization with
Lp-norm constraint. Remote Sens. 2019, 11, 559. [CrossRef]
23. Dai, Y.; Wu, Y.; Zhou, F.; Barnard, K. Attentional local contrast networks for infrared small target detection. IEEE Trans. Geosci.
Remote Sens. 2021, 59, 9813–9824. [CrossRef]
Remote Sens. 2022, 14, 4785 13 of 13

24. Zhou, F.; Wu, Y.; Dai, Y. Infrared small target detection via incorporating spatial structural prior into intrinsic tensor sparsity
regularization. Digit. Signal Process. 2021, 111, 102966. [CrossRef]
25. Tian, Y.; Liu, J.; Zhu, S.; Xu, F.; Bai, G.; Liu, C. Ship Detection in Visible Remote Sensing Image Based on Saliency Extraction and
Modified Channel Features. Remote Sens. 2022, 14, 3347. [CrossRef]
26. Deng, L.; Zhang, J.; Xu, G.; Zhu, H. Infrared small target detection via adaptive M-estimator ring top-hat transformation. Pattern
Recognit. 2021, 112, 107729. [CrossRef]
27. Hou, Q.; Wang, Z.; Tan, F.; Zhao, Y.; Zheng, H.; Zhang, W. RISTDnet: Robust infrared small target detection network. IEEE Geosci.
Remote Sens. Lett. 2021, 19, 3050828. [CrossRef]
28. Shahraki, H.; Aalaei, S.; Moradi, S. Infrared small target detection based on the dynamic particle swarm optimization. Infrared
Phys. Technol. 2021, 117, 103837. [CrossRef]
29. Zhao, M.; Li, W.; Li, L.; Ma, P.; Cai, Z.; Tao, R. Three-order tensor creation and tucker decomposition for infrared small-target
detection. IEEE Trans. Geosci. Remote Sens. 2021, 60, 5000216. [CrossRef]
30. Peršić, J.; Petrović, L.; Marković, I.; Petrović, I. Spatiotemporal multisensor calibration via gaussian processes moving target
tracking. IEEE Trans. Robot. 2021, 37, 1401–1415. [CrossRef]
31. Cui, H.; Li, L.; Liu, X.; Su, X.; Chen, F. Infrared Small Target Detection Based on Weighted Three-Layer Window Local Contrast.
IEEE Geosci. Remote Sens. Lett. 2021, 19, 3133649. [CrossRef]
32. Xue, W.; Qi, J.; Shao, G.; Xiao, Z.; Zhang, Y.; Zhong, P. Low-Rank Approximation and Multiple Sparse Constraint Modeling for
Infrared Low-Flying Fixed-Wing UAV Detection. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 4150–4166. [CrossRef]
33. Stojnić, V.; Risojević, V.; Muštra, M.; Jovanović, V.; Filipi, J.; Kezić, N.; Babić, Z. A method for detection of small moving objects in
UAV videos. Remote Sens. 2021, 13, 653. [CrossRef]
34. Aghaziyarati, S.; Moradi, S.; Talebi, H. Small infrared target detection using absolute average difference weighted by cumulative
directional derivatives. Infrared Phys. Technol. 2019, 101, 78–87. [CrossRef]
35. Pang, D.; Shan, T.; Li, W.; Ma, P.; Liu, S.; Tao, R. Infrared dim and small target detection based on greedy bilateral factorization in
image sequences. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 3394–3408. [CrossRef]
36. Dai, Y.; Wu, Y.; Zhou, F.; Barnard, K. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the
IEEE/CVF Winter Conference on Applications of Computer Vision, Waikoloa, HI, USA, 3–8 January 2021; pp. 950–959.
37. Zhou, F.; Wu, Y.; Dai, Y.; Ni, K. Robust infrared small target detection via jointly sparse constraint of l 1/2-metric and dual-graph
regularization. Remote Sens. 2020, 12, 1963. [CrossRef]
38. Liu, D.; Cao, L.; Li, Z.; Liu, T.; Che, P. Infrared Small Target Detection Based on Flux Density and Direction Diversity in Gradient
Vector Field. IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens. 2018, 2528–2554. [CrossRef]
39. Li, Y.; Zhang, Y. Robust infrared small target detection using local steering kernel reconstruction. Pattern Recognit. 2018,
77, 113–125. [CrossRef]
40. Li, C.; Chen, N.; Zhao, H.; Yu, T. Multiple-beam lidar detection technology. In Proceedings of the Seventh Asia Pacific Conference
on Optics Manufacture and 2021 International Forum of Young Scientists on Advanced Optical Manufacturing (APCOM and
YSAOM 2021), SPIE, Hong Kong, China, 13–16 August 2022; Volume 12166, pp. 1773–1778.
41. Liu, H.K.; Zhang, L.; Huang, H. Small target detection in infrared videos based on spatio-temporal tensor model. IEEE Trans.
Geosci. Remote Sens. 2020, 58, 8689–8700. [CrossRef]
42. Pang, D.; Shan, T.; Li, W.; Ma, P.; Tao, R.; Ma, Y. Facet derivative-based multidirectional edge awareness and spatial–temporal
tensor model for infrared small target detection. IEEE Trans. Geosci. Remote Sens. 2021, 60, 5001015. [CrossRef]
43. Zhang, P.; Zhang, L.; Wang, X.; Shen, F.; Pu, T.; Fei, C. Edge and corner awareness-based spatial–temporal tensor model for
infrared small-target detection. IEEE Trans. Geosci. Remote Sens. 2020, 59, 10708–10724. [CrossRef]
44. Hu, Y.; Ma, Y.; Pan, Z.; Liu, Y. Infrared Dim and Small Target Detection from Complex Scenes via Multi-Frame Spatial–Temporal
Patch-Tensor Model. Remote Sens. 2022, 14, 2234. [CrossRef]
45. Sun, Y.; Yang, J.; An, W. Infrared dim and small target detection via multiple subspace learning and spatial-temporal patch-tensor
model. IEEE Trans. Geosci. Remote Sens. 2020, 59, 3737–3752. [CrossRef]
46. Pang, D.; Ma, P.; Shan, T.; Li, W.; Tao, R.; Ma, Y.; Wang, T. STTM-SFR: Spatial–Temporal Tensor Modeling with Saliency Filter
Regularization for Infrared Small Target Detection. IEEE Trans. Geosci. Remote Sens. 2022, 60, 1–18. [CrossRef]
47. Kong, X.; Yang, C.; Cao, S.; Li, C.; Peng, Z. Infrared small target detection via nonconvex tensor fibered rank approximation.
IEEE Trans. Geosci. Remote Sens. 2021, 60, 1–21. [CrossRef]
48. Biswas, S.K.; Milanfar, P. Linear support tensor machine with LSK channels: Pedestrian detection in thermal infrared images.
IEEE Trans. Image Process. 2017, 26, 4229–4242. [CrossRef]
49. Guan, X.; Zhang, L.; Huang, S.; Peng, Z. Infrared small target detection via non-convex tensor rank surrogate joint local contrast
energy. Remote Sens. 2020, 12, 1520. [CrossRef]
50. Deng, L.; Song, J.; Xu, G.; Zhu, H. When Infrared Small Target Detection Meets Tensor Ring Decomposition: A Multiscale
Morphological Framework. IEEE Trans. Aerosp. Electron. Syst. 2022, 3162–3176. [CrossRef]