IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL.

21, 2024 3502305

A PointNet-Based CFAR Detection Method for

Radar Target Detection in Sea Clutter
Xiaolin Chen , Graduate Student Member, IEEE, Kai Liu , Member, IEEE, and Zhibo Zhang , Member, IEEE

Abstract— Radar target detection on the sea surface is chal- sea clutter as K -distribution, and Xue et al. [6] propose
lenging due to the influence of sea clutter. Traditional radar target the GLRT detector for compound-Gaussian sea clutter with
detection methods cannot model the sea clutter distributions inverse Gaussian texture (GLRT-GIG). However, they still
precisely, resulting in poor target detection performance. In this
letter, we propose a novel PointNet-based method to parallelly
have some fitting errors for estimating sea clutter distribution
detect multiple targets. We extract the global features to solve a parameters.
classification problem, that is, detecting whether there exist the Furthermore, the advent of machine-learning technology has
targets in a radar echo frame, and extract the local features to presented a new solution for radar target detection that elim-
solve a segmentation problem, that is, detecting whether it has inates the need for manual feature extraction and parameter
a target in each range cell. In addition, to implement constant tuning while providing excellent robustness and universal-
false alarm rate (CFAR) detection, we apply a statistical method
by precisely adjusting the detection threshold to keep a desired ity capabilities [7]. Unsupervised clustering algorithms [8],
probability of false alarm (PFA). Simulation results show that support vector machines (SVMs) [9], deep neural networks
the proposed method can realize the target classification with a (DNNs) [10], GoogleNet [11], ResNet [12], and PointNet
95.985% total accuracy rate when the PFA is 0.1 and achieve a [13] are widely used in face recognition, image classification,
larger detection probability under a desired PFA based on the and other aspects. They process data and extract information
IPIX radar dataset compared with the baselines.
using convolution kernels. Furthermore, PointNet can directly
Index Terms— Constant false alarm rate (CFAR), PointNet, process point cloud data and extract both global and local
radar target detection, sea clutter. features. It can also handle global classification and local
segmentation tasks, which makes it highly relevant and useful
I. I NTRODUCTION for radar target detection.
From the machine-learning perspective, target detection can
R ADAR target detection on the sea surface is a challenging
task due to the presence of sea clutter, which is similar
to the target echo signals. It causes serious interferences to
be divided into two processes, that is, classifying the whole
radar echo frame and segmenting echoes in each target cell.
target detection and is difficult to achieve accurate and robust The first process is to determine if there are any targets in
target detection [1]. the observed area, and the second process is to detect the
In traditional radar target detection algorithms, echo ampli- precise location where the target exists. Chen and Huang
tude and Doppler information are typically utilized to deter- [8] propose an unsupervised clustering-based method for
mine the presence of a target, using either incoherent or identifying rain-contaminated and low-backscatter regions in
coherent detection techniques [2]. However, these methods X-band marine radar images and verify the validity of the
usually rely on manually designed parameters, which may method. Chen et al. [9] propose an SVM-based method for
not adapt well to changes in sea environments and targets. rain detection using X-band marine radar images and verify the
To address this problem, Kelly [3] proposed a method based method’s performance. In [14], a radar target detection method
on the generalized likelihood ratio test (GLRT) where sea based on DNNs is proposed to equalize the target detection
clutter is modeled as color noise. This method can achieve problem to a binary classification problem and use extracted
constant false alarm rate (CFAR) detection by dynamically features as the decision basis within a data window. Sim-
adjusting the detection threshold based on the radar clutter ulation results demonstrate its higher detection performance
data to obtain the desired probability of false alarm (PFA) than traditional methods. In [15], a dual-activated clutter
without estimating sea clutter parameters. However, due to suppression algorithm is proposed by using a convolutional
the limitation of the simplified color noise model, it suffers neural network and its performance is verified. However, these
from weak detection performance [4]. To improve it, Shi and methods require spectrally processing of the echo data, which
Shui [5] propose the optimal K detector (OKD) by modeling can increase computational complexity. They do not process
the CFAR detection, which can sharply increase PFA in the
Manuscript received 14 November 2023; revised 15 January 2024; fluctuated clutter background [5].
accepted 1 February 2024. Date of publication 7 February 2024; date of In this letter, we propose a unified architecture based
current version 15 February 2024. This work was supported by the National
Nature Science Foundation of China under Grant U2233216 and Grant on PointNet to solve the above two radar target detection
U2033215. (Corresponding author: Zhibo Zhang.) problems. Its network inputs are the radar echo data in the sea
The authors are with the School of Electronics and Information Engineering, clutter background, and its network outputs are the results of
Beihang University, Beijing 100191, China (e-mail: cxl_email_cxl@163.com;
liuk@buaa.edu.cn; zhangzhibo94@outlook.com).
target classification and segmentation. Section II introduces a
Digital Object Identifier 10.1109/LGRS.2024.3363041 radar signal model and a persymmetric transform method for
1558-0571 © 2024 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: St. Petersburg State University. Downloaded on March 01,2024 at 16:00:10 UTC from IEEE Xplore. Restrictions apply.
3502305 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 21, 2024

data preprocessing. Section III proposes the PointNet-based Then, when zk is the echo signal of sea clutter, the received
CFAR (PointNet-CFAR) detection method. Section IV gives signal after transforming is
the simulation results based on the IPIX radar dataset, and
Section V concludes this letter. z̃k = c̃k = Tck , k = 1, 2, . . . , K . (6)
II. S YSTEM M ODEL When zk contains the target range cells, the received signal
It is assumed that the radar processes the target signal in after transforming is
the range and pulse dimensions of the received echo signal to (
determine whether there is a target on a range cell. The radar z̃k = s̃k + c̃k = Tsk + Tck , k = 1, 2, . . . , H
transmits N continuous coherent pulses in a coherent process- z̃k = c̃k = Tck , k = H + 1, H + 2, . . . , K .
ing interval (CPI). After sampling, filtering, and demodulating, (7)
the echo signals contain K continuous range cells. The
echo signal in the kth range cell can be represented by an
N -dimensional complex vector zk = [zk (1), . . . , zk (N )]T , III. P ROPOSED S CHEME
where [·]T represents the transpose. Let ck represent the A. Target Classification and Segmentation
received sea clutter echo signals, which has the covariance
matrix M. Then, the received target echo signal sk = αk p+ck , For the target classification in the process of target detection,
where p represents the target Doppler steering vector and αk is it is assumed that H̃ 0, all represents all the CUTs are sea clutter,
a deterministic unknown parameter caused by target scattering and H̃ 1, all represents there are targets in some CUTs, which
and multipath effect [16]. When all of the echo signals in the can be described by the following binary test:
K range cells are sea clutter, we have (
H̃ 0, all = H̃ 0,1 ∩ H̃ 0,2 ∩ · · · ∩ H̃ 0,K
zk = ck , k = 1, 2, . . . , K . (1) (8)
H̃ 1, all = H̃ 1,1 ∪ H̃ 1,2 ∪ · · · ∪ H̃ 1,K
And if the echo signals in H range cells out of K range
cells are(echo signals of targets, we have where K is the number of CUTs, H̃ 0,k ∈ {0, 1} represents the
zk = sk + ck , k = 1, 2, . . . , H decision result that the kth range cell is radar echo signal of
(2) sea clutter, and H̃ 1,k ∈ {0, 1} represents the decision result
zk = ck , k = H + 1, H + 2, . . . , K
that the kth range cell has radar echo signal of the target.
where zk represent the received echo signal. z1,..., z H are the For the target classification process and target segmentation
observation vectors in the cells under tests (CUTs), called process, the neural network can be regarded as a nonlinear
the primary data. z H +1 , . . . , z K are the observation vectors function f (·) and f ′ (·) of the input data, respectively. Their
in the reference cells, called the secondary data. output is
In a radar system, the clutter covariance matrix has the 

persymmetric property when a linear array is homogeneous P H0, all | Z

f (Z) = (9)
and the coherent pulse train is uniformly spaced [17]. Knowing P H1, all | Z
the prior information, the clutter covariance matrix M has a


P H0,k | Z

doubly symmetric structure f ′ (Z) = (10)
P H1,k | Z
M = JM∗ J (3)
where Z = [z1 , z2 , . . . , z K ] is the input radar data, P(H0, all |
where (·)∗ denotes the complex conjugate operator, and J is Z) is the probability that there is no target in the input data Z,
a matrix where the cross diagonal entries are ones and the that is, Z are the echo signals of pure sea clutter, P(H1, all | Z)
others are zeros is the probability that there are targets in the input data Z, and
0 0 ... 0 1
 
P(H0,k | Z) represents the probability that the kth range cell
0 0 ... 1 0 in the input data belongs to the echo signal of pure sea clutter.
J=. .. .. .. .. . (4)
 
 .. . . . .
P(H1,k | Z) represents the probability that the kth range cell
belongs to the echo signal of a target.
1 0 ... 0 0
To control the PFA, the output probability values P(H1, all |
To reduce the dependence of the detector on the reference Z) and P(H1,k | Z) can be sorted in ascending order sep-
cell data and improve the target detection probability under the arately, and the corresponding threshold value Tcla and Tseg
limited dataset, we can use the rotation matrix T to process can be selected under a given PFA, respectively. Thus, Tcla
the received signal [17], which can be expressed as and Tseg denote the classification threshold and segmentation

threshold, respectively. Specifically, assuming that the PFA
!
1 I N /2 J N /2
,

√ for N is even value is PFA , ⌊(1 − PFA )N ⌋ can be selected as the threshold,


 2  iI N /2 −iJ N /2

where ⌊·⌋ means rounded down and N indicates the number


T= I(N −1)/2 0 J(N −1)/2 of sea clutter samples or the number of sea clutter range cells
1  √
, for N is odd

√ 0 2 0 in the training set.

 
 

 2 Then, in the classification process, when the predicted value
 iI(N −1)/2 0 −iJ(N −1)/2
P(H1, all | Z) is larger than or equal to the threshold Tcla ,
(5)
it is considered that there are targets in the input data Z.
where I represents the identity matrix. On the other hand, if P(H0, all | Z) > (1 − Tcla ), that is,

Authorized licensed use limited to: St. Petersburg State University. Downloaded on March 01,2024 at 16:00:10 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: PointNet-BASED CFAR DETECTION METHOD 3502305

Then, to merge multiple features, the feature vectors of all

points are subjected to a max pooling operation MP[·] to obtain
a global feature vector Y3

Y3 = MP[Y2 ]. (15)

In the classification process, its purpose is to decide whether

there is a target in the current frame. Generally, the number
of classified categories is used as the dimension of the final
output. In this letter, the classified categories contain two
Fig. 1. Network structure of the PointNet-CFAR scheme. categories with H0, all and H1, all , that is, “no-target” and “with-
target.” Therefore, the dimension of output Y4 is 2. To reduce
P(H1, all | Z) < (Tcla ), it is considered that the target is not the influence of point cloud spatial features on classification
included in the input data. Thus, we have results and the time complexity, in the MLP-3 module, we use
( two consecutive fully connected layers f fc,m (·) to realize the
H̃ 0, all : P H0, all | Z > 1 − Tcla


(11) classification
H̃ 1, all : P H1, all | Z ≥ Tcla .


Y4 = f fc,2 σ BN4 DP f fc,1 (Y3 )
   

(16)
Similarly, in segmentation process, we compare the thresh-
old Tseg with the segmentation network output probability where DP[·] is dropout operation to prevent overfitting.
( Finally, to map the two neuron outputs represented by
H̃ 0,k : P H0,k | Z > 1 − Tseg


(12) Y4 into interval [0, 1], we use the softmax activation function
H̃ 1,k : P H1,k | Z ≥ Tseg .


to normalize and obtain the probability of the binary hypoth-
esis in the above classification process
In this way, the CFAR detection can be controlled by the
predicted value to achieve effective target detection on the 

eY4,H0, all
 P H0, all | Z = Y4,H

sea surface. The threshold can be obtained by Monte Carlo 
simulation based on the given PFA. e 0, all + eY4,H1, all (17)

eY4,H1, all
 P H1, all | Z =

eY4,H0, all + eY4,H1, all
B. PointNet-CFAR Scheme
PointNet is a DNN architecture that can directly process where Y4,H0, all and Y4,H1, all are the predicted probabilities of
point cloud data by extracting their features. It can map point whether the input data Z contain targets or not from the
cloud data of different numbers and orders into a fixed length classification network, respectively.
vector so that it can easily perform tasks such as classification, Finally, we sort the probability results obtained in (17)
segmentation, and reconstruction of point cloud data. Thus, under pure sea clutter samples and then obtain the threshold
we focus on the problem of sea surface target detection in the according to Section III-A. By substituting this threshold
background of sea clutter and apply and modify the PointNet into (11), we obtain the final classification results and achieve
network to realize two processes: target classification and the CFAR detection.
target segmentation. In the segmentation process, we aim to determine the target
As shown in Fig. 1, we use a b × K × c radar signal locations. Thus, we cascade the global features Y3 , local
tensor as the input of the network, where b is the batch size of features Y1 , and original data Z to form a new matrix Y5
parallelly training neural networks, K is the number of CUTs,
Y5 = Y1 , Ỹ 3 , Z
 
(18)
and c = 2N means the total dimension of the in-phase phase
and orthogonal phase of the N -pulse-echo signals. where Ỹ 3 = [Y3T , Y3T , . . . , Y3T ]T .
First, we obtain the local features of the MLP-1 module. We Since the above-cascaded features are high-dimensional
use multiple 1-D convolution kernels to increase the dimen- tensors, we perform 1-D convolutions on the input cascaded
sionality of the input, and at the same time, we normalize features in the MLP-4 module to obtain Y6
the convolution results to prevent overfitting. Let Fm (x) =
σ {BNm [ f conv,m (x)]}, where m is the mth layer in the network Y6 = F4 (Y5 )
structure, f conv (·) is the 1-D convolution operation, BN[·] is = σ BN5 f conv,5 σ BN4 f conv,4 (Y5 )
    

. (19)
the normalization operation, and σ {·} is the ReLu activation
function. Then, the output of MLP-1 module Y1 is The dimension of the final output depends on the number
Y1 = F1 (Z) = σ BN1 f conv,1 (Z) .
  
(13) of segmented categories, which include two parts, that is,
“clutter” and “target.” The output Y7 of the MLP-5 module is
Second, in the MLP-2 module, we use multiple 1-D convo-
lution kernels to perform two consecutive encoding operations Y7 = f conv,6 (F5 (Y6 ))
= f conv,6 σ BN5 f conv,5 (Y6 ) ⊕ Y6
  

on local features to further extract global features as (20)
Y2 = BN3 f conv,3 (F2 (Y1 ))
 
where ⊕ is the element-wise addition often used in ResNet
= BN3 f conv,3 σ BN2 f conv,2 (Y1 ) .
   

(14) for extending the network depth.

Authorized licensed use limited to: St. Petersburg State University. Downloaded on March 01,2024 at 16:00:10 UTC from IEEE Xplore. Restrictions apply.
3502305 IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 21, 2024

Similarly, to map the outputs of the two neurons represented

by Y7 to interval [0, 1] and achieve CFAR detection, we nor-
malize them using the softmax activation function to obtain
the probability of the binary hypothesis in the segmentation
eY7,H0,k



 P H0,k | Z = Y7,H


e 0,k + eY7,H1,k (21)

eY7,H1,k
 P H1,k | Z =

eY7,H0,k + eY7,H1,k
Fig. 2. Classification experiment results when all input data are predicted to
where Y7,H0,k is the predicted probability of the segmentation be the “with-target” class and Tcla = 0.5000. (a) Probability density function
network for the “clutter,” and Y7,H1,k is the predicted proba- for the “with-target” class, and (b) confusion matrix.
bility of the segmentation network for the “target.”
Similar to CFAR detection for the classification process,
we obtain the threshold Tseg under pure clutter samples. By
substituting Tseg into (12), the final detection results of the
segmentation process are obtained for CFAR detection.

IV. S IMULATION R ESULTS

In this section, we verify the target detection performance of
the proposed method by combining measured data and Monte
Carlo simulation. The sea clutter dataset comes from the IPIX-
1998 radar in Canada [18]. Fig. 3. Confusion matrices of classification experiments of different PFAs.
(a) PCFA = 0.1. (b) PCFA = 0.001.
We use part of the sea clutter of the IPIX radar in the target
detection process. Then, we model the possible targets in each
range cell according to (1) and (2) with N = 16 and the
normalized target Doppler frequency f d ∈ [−0.5, 0.5]. In the know that the threshold Tcla = 0.6567, the actual PCFA is
simulation, we use K ∈ [32, 40] and randomly select target 0.0013, and the total accuracy rate of this experiment is
cells. The signal-to-clutter ratio (SCR) is randomly selected 90.100%. When the threshold is larger than the threshold
from 5 to 10 dB. For the classification problem, we divide in Fig. 2, its corresponding PFA and total accuracy will
all the data into “no-target” and “with-target” according to decrease. Otherwise, its corresponding PFA and total accuracy
whether there exist targets in all K cells. “no-target” sets will increase. Thus, the classification performance can be
indicate that there is no target in all K cells, which contain controlled by the PFA, which is consistent with the CFAR
40% samples, and “with-target” sets indicate that there exist detection theory.
targets in all K cells, which contain 60% samples. For the In the segmentation experiment, we take the actual prob-
segmentation problem, we label each range cell according to ability of segmentation false alarms P̂ SFA and segmentation
whether it is an echo signal of the sea clutter or an echo signal detection probability P̂ Sd as the evaluation metrics of the
of the target, that is, “clutter” or “target.” experimental results. They are defined as follows:
Fig. 2(a) shows the probability density function (pdf) of Ndc
the probability that all input data are predicted to be the P̂ SFA = × 100% (22)
NTC
“with-target” class in the classification experiment. The blue NDT
curve represents the pdf of the “no-target” set predicted as the P̂ Sd = × 100% (23)
NTT
“with-target” class, and the orange curve represents the pdf
of the “with-target” set predicted as the “with-target” class. where Ndc represents the total number of detected sea clutter
In this classification process, we set Tcla = 0.5000 as the range cells, NTC represents the total number of true sea clutter
marked black solid line. If the probability of the “with-target” range cells, NDT is the total number of detected target range
class is greater than the threshold Tcla , the data is predicted cells, and NTT is the total number of true target range cells.
as the “with-target” class. Otherwise, the data is predicted Fig. 4(a) shows the CFAR performance of the seg-
as the “no-target” class. In the classification experiment, the mentation experiment. In the segmentation experiment,
probability of classification false alarm PCFA represents the we set the probability of segmentation false alarms
proportion of the number of “no-target” sets predicted as PSFA = [0.1, 0.05, 0.01, 0.005, 0.001, 0.0005]. According to
“with-target” class in the input data to the total number of the CFAR detection process proposed in Section III-A, we con-
“no-target” set. Fig. 2(b) shows the confusion matrix of this duct the segmentation experiment with CFAR performance. In
experiment. We can know that the actual PCFA is 0.0047 and the figure, the yellow dotted line represents the given PFA,
the total accuracy rate of this experiment is 99.726%. and the blue solid line represents the actual PFA output in
Fig. 3(a) shows the confusion matrix of classification results the segmentation experiment. We can see that the actual PFA
when PCFA = 0.1. We can know that the threshold Tcla = almost coincides with the given PFA.
0.4590, the actual PCFA is 0.1001, and the total accuracy rate Fig. 4(b) shows the locations of false alarm range cells
of this experiment is 95.985%. Fig. 3(b) shows the confusion obtained from a set of IPIX-1998 radar sea clutter data after
matrix of classification results when PCFA = 0.001. We can the segmentation experiment when PSFA = 0.001. There are

Authorized licensed use limited to: St. Petersburg State University. Downloaded on March 01,2024 at 16:00:10 UTC from IEEE Xplore. Restrictions apply.
CHEN et al.: PointNet-BASED CFAR DETECTION METHOD 3502305

Fig. 4. Segmentation results. (a) CFAR performance with different PFAs. (b) Locations of false alarm range cells on the IPIX radar dataset. (c) Detection
probability.

101 250 CUTs in total, where 56 307 CUTs are echo signals [2] C. Kuang, C. Wang, B. Wen, Y. Hou, and Y. Lai, “An improved CA-
of pure sea clutter. When PSFA = 0.001, this group of data CFAR method for ship target detection in strong clutter using UHF
radar,” IEEE Signal Process. Lett., vol. 27, pp. 1445–1449, Aug. 2020,
should theoretically produce 56 false alarm range cells. In doi: 10.1109/LSP.2020.3015682.
the experiment, the actual PFA is 0.00094, and this indicates [3] E. J. Kelly, “An adaptive detection algorithm,” IEEE Trans. Aerosp. Elec-
53 false alarm range cells. The white dots mark the locations tron. Syst., vol. AES-22, no. 2, pp. 115–127, Mar. 1986.
[4] S. Xu, X. Shi, J. Xue, and P. Shui, “Adaptive subspace detection of
of actual false alarm range cells in the figure. range-spread target in compound Gaussian clutter with inverse Gaussian
Fig. 4(c) compares the detection results of the proposed texture,” Digit. Signal Process., vol. 81, pp. 79–89, Oct. 2018.
PointNet-CFAR method, the PointNet-CFAR method without [5] S. Shi and P. Shui, “Optimum coherent detection in homogenous
the ResNet part, and the cascaded original input data, DNN K-distributed clutter,” IET Radar, Sonar Navigat., vol. 10, no. 8,
pp. 1477–1484, Oct. 2016.
with CFAR detection, and GLRT detector [3]. It shows that [6] J. Xue, S. Xu, and P. Shui, “Near-optimum coherent CFAR detection
the proposed PointNet-CFAR method has better detection of radar targets in compound-Gaussian clutter with inverse Gaussian
performance than the other methods because of the full use of texture,” Signal Process., vol. 166, Jan. 2020, Art. no. 107236.
the global and local features of radar echo signals. In addition, [7] C. Wang, J. Tian, J. Cao, and X. Wang, “Deep learning-based UAV
detection in pulse-Doppler radar,” IEEE Trans. Geosci. Remote Sens.,
it is reasonable that the detection probability increases with the vol. 60, 2022, Art. no. 3104907.
increase of the given PFA. This is because the increased PFA [8] X. Chen and W. Huang, “Identification of rain and low-backscatter
causes the lower calculated threshold, which means it is easy regions in X-band marine radar images: An unsupervised approach,”
IEEE Trans. Geosci. Remote Sens., vol. 58, no. 6, pp. 4225–4236,
to exceed the threshold and the probability of the target being Jun. 2020.
detected is higher. Furthermore, we can see that the DNN [9] X. Chen, W. Huang, C. Zhao, and Y. Tian, “Rain detection from X-band
method based on machine learning outperforms the traditional marine radar images: A support vector machine-based approach,” IEEE
Trans. Geosci. Remote Sens., vol. 58, no. 3, pp. 2115–2123, Mar. 2020.
GLRT detector in the effective detection interval. However,
[10] M. Lin, W. Wang, X. Hong, and W. Zhang, “GLRT approach for multi-
it only has a higher detection probability in the high PFA antenna-based spectrum sensing under interference,” IEEE Commun.
interval, and always has a lower detection probability than the Lett., vol. 24, no. 7, pp. 1524–1528, Jul. 2020.
proposed PointNet-CFAR method. [11] Q. Wang, W. Huang, Z. Xiong, and X. Li, “Looking closer at the
scene: Multiscale representation learning for remote sensing image scene
classification,” IEEE Trans. Neural Netw. Learn. Syst., vol. 33, no. 4,
V. C ONCLUSION pp. 1414–1428, Dec. 2022.
In this letter, a PointNet-CFAR detection method is pro- [12] J. He, S. Liang, X. Wu, B. Zhao, and L. Zhang, “MGSeg: Multi-
posed for the background of sea clutter to overcome the poor ple granularity-based real-time semantic segmentation network,” IEEE
Trans. Image Process., vol. 30, pp. 7200–7214, 2021.
detection performance of traditional methods. The method [13] R. Q. Charles, H. Su, M. Kaichun, and L. J. Guibas, “PointNet:
directly takes the radar echo signal as the input and provides Deep learning on point sets for 3D classification and segmentation,”
a unified architecture to solve the problem of target classi- in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), Jul. 2017,
pp. 77–85.
fication and segmentation by extracting global features and [14] L. Wang, J. Tang, and Q. Liao, “A study on radar target detection based
local features. It can detect multiple targets in echo signals in on deep neural networks,” IEEE Sensors Lett., vol. 3, no. 3, pp. 1–4,
parallel and realize CFAR detection by controlling the decision Mar. 2019.
threshold at the same time. Simulation results show that the [15] G.-Q. Li, Z.-Y. Song, and Q. Fu, “A convolutional neural network
based approach to sea clutter suppression for small boat detection,”
detection performance of the proposed method is better than Frontiers Inf. Technol. Electron. Eng., vol. 21, no. 10, pp. 1504–1520,
the baselines. In the future, we will study the collaborative Oct. 2020.
detection of multitime echoes and model the rain clutter to [16] Z. Wang, Z. He, Q. He, and Z. Cheng, “Persymmetric range-spread
targets detection in compound Gaussian sea clutter with inverse Gaus-
further improve the detection performance. sian texture,” IEEE Geosci. Remote Sens. Lett., vol. 19, pp. 1–5,
2022.
[17] Z. Wang, G. Li, and H. Chen, “Adaptive persymmetric subspace detec-
R EFERENCES tors in the partially homogeneous environment,” IEEE Trans. Signal
[1] H. Zhou and T. Jiang, “Decision tree based sea-surface weak target Process., vol. 68, pp. 5178–5187, 2020.
detection with false alarm rate controllable,” IEEE Signal Process. Lett., [18] IPIX Radar Database. Hamilton, ON, Canada: McMaster Univ.,
vol. 26, no. 6, pp. 793–797, Jun. 2019. 2012.

Authorized licensed use limited to: St. Petersburg State University. Downloaded on March 01,2024 at 16:00:10 UTC from IEEE Xplore. Restrictions apply.