sensors

Article
LPDNet: A Lightweight Network for SAR Ship Detection Based on
Multi-Level Laplacian Denoising
Congxia Zhao 1 , Xiongjun Fu 1,2, *, Jian Dong 1 , Cheng Feng 1 and Hao Chang 1

1 Beijing Institute of Technology, Beijing 100081, China; zcx09890989@163.com (C.Z.);

radarvincent@sina.com (J.D.); dafeng7@126.com (C.F.); 3120200660@bit.edu.cn (H.C.)
2 Tangshan Research Institute of BIT, Tangshan 063000, China
* Correspondence: fuxiongjun@bit.edu.cn

Abstract: Intelligent ship detection based on synthetic aperture radar (SAR) is vital in maritime
situational awareness. Deep learning methods have great advantages in SAR ship detection. However,
the methods do not strike a balance between lightweight and accuracy. In this article, we propose an
end-to-end lightweight SAR target detection algorithm, multi-level Laplacian pyramid denoising
network (LPDNet). Firstly, an intelligent denoising method based on the multi-level Laplacian
transform is proposed. Through Convolutional Neural Network (CNN)-based threshold suppression,
the denoising becomes adaptive to every SAR image via back-propagation and makes the denoising
processing supervised. Secondly, channel modeling is proposed to combine the spatial domain and
frequency domain information. Multi-dimensional information enhances the detection effect. Thirdly,
the Convolutional Block Attention Module (CBAM) is introduced into the feature fusion module
of the basic framework (Yolox-tiny) so that different weights are given to each pixel of the feature
map to highlight the effective features. Experiments on SSDD and AIR SARShip-1.0 demonstrate
that the proposed method achieves 97.14% AP with a speed of 24.68FPS and 92.19% AP with a speed
of 23.42FPS, respectively, with only 5.1 M parameters, which verifies the accuracy, efficiency, and
lightweight of the proposed method.

Keywords: synthetic aperture radar (SAR); Convolutional Neural Network (CNN); target detection;
trainable; Laplacian pyramid denoising
Citation: Zhao, C.; Fu, X.; Dong, J.;
Feng, C.; Chang, H. LPDNet: A
Lightweight Network for SAR Ship
Detection Based on Multi-Level
1. Introduction
Laplacian Denoising. Sensors 2023, 23,
6084. https://doi.org/10.3390/
Synthetic aperture radar (SAR) is an active microwave sensor with strong penetrability,
s23136084
which produces high-resolution images at all-time and all-weather. It is widely used in
disaster prevention, emergency rescue, land resources monitoring, urban management,
Academic Editor: Gwanggil Jeon
etc. [1,2]. Ship detection based on SAR images effectively monitors ocean conditions and
Received: 25 May 2023 manages ocean resources [3].
Revised: 17 June 2023 Traditional SAR ship detection methods are mainly based on the statistical features [4–6],
Accepted: 20 June 2023 scattering characteristics [7], and transform domain [8], while the most typical method
Published: 1 July 2023 is the constant false alarm rate (CFAR) [9–12]. However, its feature extraction process is
complex, and it is sensitive to speckle noise and complex backgrounds, which can not meet
the requirements of high-precision, lightweight, and real-time SAR ship detection.
With the emergence and vigorous development of CNNs, object detection algorithms
Copyright: © 2023 by the authors. based on deep learning have become the mainstream methods of image processing. The
Licensee MDPI, Basel, Switzerland. algorithms fall into two main categories: one-stage [13–16] and two-stage [17–20] meth-
This article is an open access article
ods. The one-stage algorithms, including Yolo [13], SSD [14], RetinaNet [15], etc., produce
distributed under the terms and
the detection results directly from the figures. The two-stage algorithms, such as Faster
conditions of the Creative Commons
R-CNN [18], Mask R-CNN [19], and Cascade R-CNN [20], divide the target detection into
Attribution (CC BY) license (https://
two stages, generating region proposal and classification. In recent years, with the proposal
creativecommons.org/licenses/by/
of SAR datasets, containing SSDD [21], HRSID [22], and OpenSARship [23], SAR target
4.0/).

Sensors 2023, 23, 6084. https://doi.org/10.3390/s23136084 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 6084 2 of 16

detection based on deep learning has developed towards high precision, high speed, and
lightweight. The two-stage methods generally possess higher accuracy. Y. Li et al. [24]
designed a feature relay amplification and multiscale feature jump connection structure.
They proposed a lightweight network based on Faster R-CNN [18] to improve the detection
speed while ensuring detection accuracy. Cui et al. [25] densely connected CBAM [26] to
each concatenated feature map from top to bottom of the pyramid network to obtain abun-
dant features containing resolution and semantic information and improve the detection
accuracy of multiscale ships. The one-stage detectors reduce the amount of calculation
through various methods to improve detection efficiency. H. Wan et al. [27] designed a
lightweight backbone network based on Yolox [28], highlighting SAR targets’ unique strong
scattering characteristics by integrating channels and spatial attention mechanisms and
improving detection accuracy and speed. Qi et al. [29] introduced weak segmentation and
attention mechanism [30] into a single-stage detector to obtain richer semantic informa-
tion. Mao et al. [31] proposed a simple detector based on U-Net [32], but the detection
performance was unsatisfactory. These methods are competitive in terms of accuracy
and lightweight. However, they process SAR images as optical images and ignore their
differences. SAR images are produced in microwave/millimeter wave bands, which are
different from common three-channel images, RGB or HSV, acquired by visible and partial
infrared band sensors. The differences make the CNN-based methods not fully applicable
to SAR image processing. Moreover, because of the basic principle of coherent imaging,
SAR images contain a lot of speckle noise, which brings a low signal-to-noise ratio.
In SAR images, speckle noise can easily lead to false detection and target missing.
Traditional image denoising mainly includes spatial domain-based methods and frequency
domain-based methods. L. Liu et al. [33] used morphological filtering for SAR image
denoising to preprocess image change detection. R. Farhadiani et al. [34] used a maximum
a posteriori (MAP) estimator to denoise in the complex wavelet domain and utilized the
local pixel group filtering based on the non-local principal component analysis (LPG-PCA)
method to smooth the homogeneous areas and enhance the details. Through feature
extraction of CNNs, SAR image denoising methods based on deep learning extract the
effective features and obtain the noise and details in the image. J. Zhang et al. [35] performed
SAR image denoising by a multiconnection network that incorporates wavelet features
while maintaining the texture structure of the image. S. Liu et al. [36] extracted shallow
features from noisy images by designing kernels of different sizes to form multiscale
modules. Following this, the shallow features are mapped to the residual dense dual-
attention network to obtain the deep features of SAR images, and the final denoised images
are generated through global residual learning. However, most of the above methods only
regard SAR image denoising as the preprocessing and do not realize the end-to-end image
processing. At the same time, in the traditional denoising methods, the noise suppression
thresholds are often set according to experience, which is vulnerable to subjective influence,
and the self-adaptation to each image, each network, and target detection task cannot
be achieved.
This article proposes an integrated model of the transform domain denoising method
and target detection network, LPDNet, to solve the above problems. The main contributions
can be summarized as follows:
(1) To reduce speckle noise and improve detection performance, this article proposes a
novel intelligent denoising method based on Laplacian pyramid denoising and CNN.
Through CNN-based threshold suppression, the multi-level Laplacian denoising
becomes adaptive to every SAR image via back-propagation based on the subsequent
detection network, which makes the denoising processing supervised.
(2) This article proposes to perform channel modeling on feature maps composed of
spatial sub-bands, denoised maps, and original maps. It strengthens the balance
between the contributions of spatial and frequency features, targets, and noise, thereby
effectively combining the multi-dimensional information.
Sensors 2023, 23, 6084 3 of 16

(3) In order to further improve the accuracy of the model, the CBAM is introduced to the
feature fusion module of the basic framework. The CBAM is flexible and lightweight,
which avoids a lot of computational overhead, and the accuracy is compensated.
(4) In this article, supervised thinking in deep learning is combined with traditional
image decomposition and enhancement methods. It provides a new supervised
denoising method for SAR target detection. The method can be embedded in SAR
image processing as a preprocessing process.
The feasibility and effectiveness of LPDNet are verified on the SAR ship detection data
set (SSDD), AIR SARShip-1.0 [37], and HRSID. The method is compared with other CNN-
based algorithms and evaluated by indicators, such as average precision mean (mAP),
precision, recall, F1-score, frame per second (FPS), and parameters. The experimental
results prove its accuracy and efficiency.
The remainder of this article is organized as follows: Section 2 introduces our proposed
target detection method, Section 3 presents experiments and results analysis, and Section 4
concludes the paper.

2. Methodology
In this section, the multi-level Laplacian pyramid denoising network is developed,
and details of the implementation procedures are presented. First, an overview of LPDNet
is introduced. Then, the structure of multi-level Laplacian pyramid denoising is developed.
Finally, CBAM is introduced into the model to enhance the representation ability.

2.1. Processing Flow of LPDNet

The architecture of the proposed network, LPDNet, is shown in Figure 1. LPDNet
firstly derives the Laplacian pyramid intelligent denoising module to reduce speckle noise.
The multi-level Laplacian pyramid transformer is applied to decompose the image into
different frequency sub-bands. A module based on CNN is used to intelligently determine
the threshold of each high-frequency sub-band through back-propagation. The process
of back-propagation in Figure 1 is depicted as a dotted arrow. The adaptive thresholds
are generated and extracted through the forward propagation and adjusted through the
back-propagation so that the thresholds can be adaptively determined by the detection
task. Then, hard thresholding is used to suppress the noise. The denoised image can
be obtained through Laplacian pyramid reconstruction. To compensate for the feature
lost in the denoising and enhance the high-frequency information, channel modeling is
performed. The three-channel image consists of a high-frequency image of the first level
of the Laplacian pyramid decomposition, a denoised image, and an original image. After
that, CBAM is introduced into the feature fusion part of the Yolox-tiny to enhance feature
extraction. The details of the method are shown as follows.

2.2. Multi-Level Laplacian Pyramid Denoising Model

Laplacian pyramid transform is a multiscale and multiresolution image processing
method that can decompose an image’s high-frequency and low-frequency features into
sub-bands of different resolutions according to different scales. Speckle noise in SAR images
mainly distributes in the high-frequency sub-bands of the image, with relatively small
coefficients. By threshold suppression, the coefficients that are lower than the threshold
can be compressed. Then, the denoised image can be obtained by Laplacian pyramid
reconstruction. However, the traditional threshold determination method is usually affected
by subjective factors, which cannot achieve the self-adaptation of each image, each network,
and target detection task. In this section, a CNN-based Laplacian pyramid denoising
method is proposed, which converts the threshold into a trainable variable through the
network. The denoising threshold of each high-frequency sub-band can be determined
intelligently through the supervised learning of the subsequent detection network.
 ing is used to suppress the noise. The denoised image can be obtained through Laplacian pyr-
amid reconstruction. To compensate for the feature lost in the denoising and enhance the high-
frequency information, channel modeling is performed. The three-channel image consists of
a high-frequency image of the first level of the Laplacian pyramid decomposition, a denoised
Sensors 2023, 23, 6084 image, and an original image. After that, CBAM is introduced into the feature fusion part4 of of
16
the Yolox-tiny to enhance feature extraction. The details of the method are shown as follows.

Back-propagation

Multi-level Laplacian Pyramid Denoising Model

IH1 CBAM Yolox-tiny

TH

IH2 LPR
LPD C
TH

LPR
LPD IH3
TH
I L1
LPD LPR
C C
I L2
I L3

C C
LPD: Laplacian Pyramid Decomposition TH: CNN-based threshold denoising
LPB: Laplacian Pyramid Reconstruction C : CBAM C : Concat Yolo Head Yolo Head Yolo Head

Figure 1. Processing flow of the proposed network.

2.2.1. Laplacian Pyramid Denoising Model

The Laplacian pyramid denoising model is shown in Figure 2, where LPD is the
Laplacian pyramid decomposition, and LPR is the Laplacian pyramid reconstruction.
The high-frequency and low-frequency sub-bands are obtained by Laplacian pyramid
decomposition, and the high-frequency sub-band is put into the threshold suppression
module for noise suppression. The CNN-based threshold suppression module consists of a
convolution layer, a fully connected layer, and a hard threshold function. The convolution
kernel size of the convolution layer is the same as the size of the input high-frequency
sub-band image and has the global receptive field of the image, which can be expressed as:

Oi = MaxPool ( f i ∗ IH + bi ) (1)

where Oi is the ith feature channel of the convolutional layer output, MaxPool () is the
maximum pooling function, f i represents the convolution kernel of the input feature
corresponding to the ith channel, IH is the high-frequency sub-band image input to the
convolution layer, and bi represents the biasing of the convolutional layer. The fully
connected layer takes the output of the convolutional layer, Oi , as the input. The output (D)
is a tensor, D ∈ R1×1×1 :
D = ReLU (WO + b) (2)
where ReLU () is the relu activation function, which can be defined as follows:

x, x > 0
ReLU ( x ) = (3)
0, x ≤ 0

Then, the denoising threshold corresponding to a high-frequency sub-band can be

obtained. The noise in the high-frequency sub-band is suppressed by a hard threshold
function as follows:

0.001 × IH ( x, y), IH ( x, y) ≤ D
ÎH ( x, y) = hard_threshold( IH ( x, y)) = (4)
IH ( x, y), IH ( x, y) > D

where IH ( x, y) and ÎH ( x, y) represent the high-frequency sub-band coefficients before and
after the threshold suppression, respectively.
Finally, the Laplacian pyramid reconstruction is encapsulated into an image reconstruc-
tion layer. The low-frequency sub-band image and the denoised high-frequency sub-band
image are put into the layer to obtain the denoised image.
Sensors 2023, 23, x FOR PEER REVIEW
Sensors 2023, 23, 6084 5 of 16

Input Image
LPD

Low-frequency High-frequency CNN-based Threshold

Sub-band Sub-band Suppression

1*1 Conv

Fully
Connected
Hard
Threshold
Denoised High-
frequency Sub-band

LPR LPD:Laplacian Pyramid Decomposition

Denoised Image LPR:Laplacian Pyramid Reconstruction

Figure 2. Laplacian pyramid denoising model.

FigureThe2. Laplacian
Laplacian pyramid
denoising denoising
module based onmodel.
the CNN can be embedded in any image
processing network as the preprocessing. The convolution layer and the fully connected
layer can be trained by the loss function through the back-propagation. The denois-
Finally, the Laplacian pyramid reconstruction is encapsulated into an
ing threshold can be calculated by the global characteristics of each image to achieve
struction layer. The low-frequency sub-band image and the denoised high-f
intelligent denoising.
band image are put into the layer to obtain the denoised image.
2.2.2. Multi-Level Laplacian Pyramid Denoising Model
The Laplacian denoising module based on the CNN can be embedded
A multi-level Laplacian pyramid denoising model is designed based on the single-level
processing network
denoising structure as the
proposed preprocessing.
in Section The
2.2.1. The model convolution
contains layer
three layers, and the fu
as shown
layer can be trained by the loss function through the
in the green box in Figure 1. Firstly, the high-frequency sub-band I H1 back-propagation. T
and low-frequency
sub-band IL1 are obtained through the first layer of Laplacian pyramid decomposition.
threshold can be calculated by the global characteristics of each image to a
By decomposing the first and second low-frequency sub-bands, IH2 , IL2 and IH3 , IL3 can
gent denoising.
be obtained, respectively. Secondly, the CNN-based threshold suppression module is
constructed for each level’s high-frequency sub-bands to obtain the sub-bands after noise
suppression.
2.2.2. Then, the denoised
Multi-Level Laplacianimages are obtained
Pyramid by Laplacian
Denoising pyramid reconstruction,
Model
as shown in Figure 3. The images in the first row are the originals in the SSDD, and the
Athe
ones in multi-level
second row Laplacian pyramiddenoised
are the corresponding denoising model
images. is designed
It is obvious that the based
speckle noise in the denoised images is weaker than that in the original
level denoising structure proposed in Section 2.2.1. The model contains th images. However,
the hard threshold function may cause image distortion, including the ringing effect and
shown in the
pseudo-Gibbs green box in Figure 1. Firstly, the high-frequency sub-band
phenomenon.
frequency sub-band I L1 are obtained through the first layer of Laplacian
2.2.3. CBAM Yolox-Tiny
composition. By decomposing
The Yolox networks, theaccurate
which are more first and second
and faster thanlow-frequency
other detection al-sub-ba
gorithms, are derived from Yolo v3. Considering the accuracy, speed, and lightweight
and I H 3 , Iof
requirements
can be obtained, respectively. Secondly, the CNN-based thres
L 3 ship detection, this article uses Yolox-tiny as the basic network.

sion module is constructed for each level’s high-frequency sub-bands to o

bands after noise suppression. Then, the denoised images are obtained by L
amid reconstruction, as shown in Figure 3. The images in the first row are th
the SSDD, and the ones in the second row are the corresponding denoised
obvious that the speckle noise in the denoised images is weaker than that i
Sensors 2023, 23,
Sensors 2023, 23, 6084
x FOR PEER REVIEW 6 6of
of 16
16
(a)

(a)
(b)
Figure 3. Original images and denoised images in the SSDD; (a) original images, (b) denoised im
ages.

2.2.3. CBAM Yolox-Tiny

The Yolox networks, which are more accurate and faster than other detection alg
rithms, are derived from Yolo v3. Considering the accuracy, speed, and lightweight
quirements of ship detection,(b)this article uses Yolox-tiny as the basic network.
CBAM proposed by Woo et al. [35] contains channel and spatial attention, which d
Figure 3. Original
Figure 3. images and
Original images anddenoised
denoisedimages
imagesininthe
theSSDD;
SSDD;(a)(a)original
original images,
images, (b)(b) denoised
denoised im-
images.
ducesages.
the attention maps along the channel and spatial dimension, as shown in Figure
Channel CBAMattention is obtained
proposed by Wooby utilizing
et al. the relations
[35] contains channel and of the features
spatial between
attention, which channe
2.2.3.
deducesCBAM
Specifically, the Yolox-Tiny
theattention
global maps
information
along the is obtained
channel by two
and spatial kinds of
dimension, as pooling: maximum
shown in Figure 4. po
Channel
ing and average attention is
The Yoloxpooling. obtained
networks, which
The by utilizing the relations
are morecharacteristic
nonlinear of
accurate and faster the features
changethanis between
other channels.
detectionthrough
performed algo- a fu
Specifically,
rithms, are the global
derived information
from Yolo v3. is obtained bythe
Considering twoaccuracy,
kinds of pooling:
speed, maximum
and pooling
lightweight re-
connected network. Then, the attention weight is obtained by summing and activating
and averageofpooling.
quirements The nonlinear
ship detection, characteristic
this article change
uses Yolox-tiny is performed
as the through a fully
basic network.
two channels.
connected The
CBAMnetwork.
process
proposed Then, can be expressed
the attention
by Woo et al. [35]weight
contains
as follows:
is obtained
channel and by summing and activating
spatial attention, the
which de-
two channels. The process
duces the attention maps ( can
F ) =be
M calong σ (expressed
the MLP as follows:
( AvgPool
channel and ( F )) + dimension,
spatial MLP(MaxPool ( F ))) in Figure 4.
as shown
Channel attention is obtained by utilizing the relations of the features between channels.
Mc ( F ) = σ ( MLP( AvgPool ( F )) + MLP( MaxPool ( F ))) (5)
whereSpecifically, the global
F is the original information
feature map,is obtained
σ is theby two kinds
sigmoid of pooling:
function, andmaximum pool-
MLP indicates mu
ing and
layerwhere average
perceptron. pooling.
Similarly,
F is the original The nonlinear
two
feature map,maps characteristic
cansigmoid
σ is the change
be obtained is
function, performed
byand
using through
MLP maximum a fully
pooling a
indicates multi-
connected
layer network.
perceptron. Then, thetwo
Similarly, attention
maps weight
can be is obtained
obtained byby summing
using and activating
maximum pooling the
and
average pooling on the channel dimensions. The spatial attention weight can be obtain
two channels.
average pooling The
onprocess can be
the channel expressed as
dimensions. follows:
The spatial attention weight can be obtained
by concatenating two maps and the convolution operation.
by concatenating two maps ) = σthe
M c ( Fand (MLP + MLP(MaxPool ( F )))
( AvgPool ( F ))operation.
convolution (5)
M s ( F ) =7σ
×σ
( f 7×7 ([ AvgPool ( F ); MaxPool ( F )]))
7 is the sigmoid function, and MLP indicates multi-
where F is the originalMfeature
s ( F ) = map,
σ ( f ([ AvgPool ( F ); MaxPool ( F )])) (6)
layer
Here,perceptron. Similarly, two
7 ×maps can be obtained by using maximum pooling and
f 7×7 7×
represents the 7 convolution process.
Here,
average f 7 represents
pooling on the channel × 7 convolution
the 7 dimensions. The process.
spatial attention weight can be obtained
by concatenating two maps and the convolution operation.
Channel Attention Spatial Attention
M s ( F ) = σ ( f 7×7 ([ AvgPool ( F ); MaxPool ( F )])) (6)
MLP Max Conv
Max Here, f 7×7 represents the 7 × 7 convolution process.
C
Avg
Avg
Channel Attention Spatial Attention
F F”
MLP Max Conv
Max
C

Avg
× Avg ×
F F”

Figure 4. Structure
Figure of of
4. Structure CBAM. ×
CBAM. ×

Figure 4. Structure of CBAM.

Sensors 2023, 23, x FOR PEER REVIEW 7 of 16

Sensors 2023, 23, 6084 7 of 16

By multiplying the attention maps with the feature image and adapting the features,
By multiplying
the essential featuresthe areattention
focused,maps
andwith the featurefeatures
unnecessary image and areadapting the features,
suppressed, so that the
the essential features are focused,
representation ability is enhanced. and unnecessary features are suppressed, so that the
representation ability is enhanced.
CBAM is introduced into the feature fusion layer of the Yolox-tiny network, as shown
CBAM is introduced into the feature fusion layer of the Yolox-tiny network, as shown
in Figure 5. It is placed on the two branches of the network. Different weights are given to
in Figure 5. It is placed on the two branches of the network. Different weights are given to
each channel
each channelandandpixel ofthe
pixel of thefeature
feature map,
map, which
which provides
provides adequate
adequate information
information for fea-
for feature
ture extraction.
extraction. AtAtthethe same
same time,
time, thethe invalid
invalid information
information is suppressed,
is suppressed, thus thus the detection
the detection
accuracy is further improved.
accuracy is further improved.

CSP2_1 head
CBS = Conv BN SiLU

CSP2_1 CBS CBAM C

CSP1_n = CBS Res
Unit
C CBS
SPP Upsample CBAM Repeat n
CBS CBS
C CBS
CSP1_3 CSP2_1 head CSP2_n = CBS CBS CBS C CBS
CSP2_1
CBS Repeat n
CBS CBAM C CBS
Upsample
CSP1_3 CBAM
maxpool
CBS
C CBS maxpool
SPP = CBS C CBS
maxpool
CSP1_1
CBS CSP2_1 head
slice
slice
Focus = slice C CBS
Focus
slice
C ：concat
Input

Figure 5. CBAM
Figure 5. CBAMYolox-tiny
Yolox-tiny structure.
structure.

3. Experiments
3. Experiments
In this section, the experiments are discussed and analyzed to verify the effect of the
In this section,
proposed the experiments
method through areon
experiments discussed
the SSDD,and
AIRanalyzed to verify
SARShip-1.0 the effect
and HRSID. Theof the
proposed
experiments discuss the effectiveness of each part of the network and the comparative The
method through experiments on the SSDD, AIR SARShip-1.0 and HRSID.
experiments discuss
effect with other the effectiveness
networks. of each
All models are part of using
implemented the network and the comparative
Pytorch framework under
effect with
Linux otherwith
system networks.
NVIDIAAll models
Tesla are implemented
K80 GPUs support. using Pytorch framework under
Linux system with NVIDIA Tesla K80 GPUs support.
3.1. Dataset
To verify the model’s effectiveness, experiments are carried out on SSDD, AIR SARShip-
3.1. Dataset
1.0, and HRSID. The SSDD dataset contains 1160 SAR images, each with a size of about
500To×verify the model’s
500. These effectiveness,
images have resolutionsexperiments
of 1–15 m, witharedifferent
carried out on SSDD,
polarization AIR SAR-
modes,
Ship-1.0, and HRSID. The SSDD dataset contains 1160 SAR images, each
including HH, HV, VH, and VV. There are 2358 ships with different scales and materials. with a size of
about
The 500
AIR × 500. Thesecontains
SARShip-1.0 images31have resolutions
images of 1–15
of large scenes withm,a size
withofdifferent polarization
3000 × 3000 and
modes,
nearlyincluding
a thousandHH, HV,
ships VH, andtypes,
of different VV. There are 2358
with image ships with
resolutions of 1 mdifferent scales
and 3 m and a and
single polarization
materials. mode. The HRSID
The AIR SARShip-1.0 includes
contains 31 5604 images
images and 16,951
of large scenes ships
withwith different
a size of 3000 ×
polarization
3000 and nearly modes and a ranging
a thousand ships ofresolution
differentfromtypes,1m to 5image
with m. Inresolutions
the experiment,
of 1 mweand 3
randomly resize the images into 416 × 416, 416 × 416, and 800 × 800, respectively.
m and a single polarization mode. The HRSID includes 5604 images and 16,951 ships with All
datasets contain sea surfaces with different sea conditions and complex scenes, including
different polarization modes and a ranging resolution from 1 m to 5 m. In the experiment,
inshore and offshore. For the convenience of the experiment, the data sets are divided into
wearandomly resize
train set, a test set,the
andimages into set
a validation 416according
× 416, 416to × 416,
the and
ratio 800 × 800, respectively. All
of 7:2:1.
datasets contain sea surfaces with different sea conditions and complex scenes, including
inshore and offshore. For the convenience of the experiment, the data sets are divided into
a train set, a test set, and a validation set according to the ratio of 7:2:1.
Sensors 2023, 23, 6084 8 of 16

3.2. Evaluation Metrics

In the experiments, precision, recall, average precision (AP), and F1 are employed to
evaluate the performance of the methods. Precision and recall can be described as follows:

TP
Precision = (7)
TP + FP

TP
Recall = (8)
TP + FN
where TP means the ships are detected correctly, FP means missed detection, and FN means
false alarm.
AP and F1 measure the balance between precision and recall:
Z 1
AP = P( R)· RdR (9)
0

2· R · P
F1 = (10)
R+P
where P respects precision and R respects recall, AP computes the average value of precision
over the interval from recall = 0 to 1, and F1 is the harmonic mean of precision and recall.
Both of them are comprehensive calculation indicators of precision and recall, and the
higher the value, the better the detection effect.

3.3. Evaluation of the Methods with Different Thresholds

In this section, the effect of the intelligent Laplacian denoising in the LPDNet is
compared with the methods based on fixed thresholds to verify the effectiveness of the
CNN-based threshold determination method.
From the network combining Yolox-tiny with the multi-level Laplacian pyramid
denoising model (LP Yolox-tiny), the value ranges of the three intelligently determined
thresholds are [0, 1.143], [0, 0.579], and [0, 0.881], respectively. We take the medians, means,
maximums, and minimums of the ranges as the thresholds of the experiments, respectively.
The rest of the module settings are the same as the LP Yolox-tiny. As shown in Table 1, the
effects of methods with fixed thresholds are inferior to that of the CNN-based method (LP
Yolox-tiny). In particular, the AP of LP Yolox-tiny achieves 96.25%, which is 0.51%, 0.5%,
1.43%, and 1.28% higher than that of the methods taking medians, means, maximums, and
minimums as thresholds, respectively. Compared with the fixed threshold, the intelligently
determined threshold makes the denoising effect more suitable for the detection network,
so the detection effect is advantageous. The advantages of the CNN-based method can be
further reflected by the precision-recall (PR) curve, as shown in Figure 6.

Table 1. The results of experiments with different thresholds.

Methods Threshold1 Threshold2 Threshold3 Recall Precision AP F1

Median 0.024 0 0.056 89.01% 96.05% 95.74% 0.92
Mean 0.087 0.024 0.083 89.01% 94.55% 95.75% 0.92
Maximum 1.143 0.579 0.881 87.55% 94.09% 94.82% 0.91
Minimum 0 0 0 89.01% 94.37% 94.97% 0.92
CNN-based -- -- -- 91.03% 94.67% 96.25% 0.93
Sensors 2023, 23, x FOR PEER REVIEW 9 of 16
Sensors 2023, 23, 6084 9 of 16

Figure 6. PR curves of methods with different thresholds.

Figure 6. PR curves of methods with different thresholds.
3.4. Ablation Experiment
3.4. Ablation Experiment
In this section, the effect of the Laplacian denoising and CBAM in the LPDNet are
In thisbased
evaluated section, theSSDD
on the effectfrom
of the Laplacian
visual denoising
and indicator and CBAM in the LPDNet are
aspects.
evaluated based on the SSDD from visual and indicator aspects.
3.4.1. Analysis of Visualization Effects
3.4.1. To observe
Analysis of the feature extraction
Visualization Effects effect of the network, heatmaps are adopted to
visualize the effect of the network. As shown in Figure 7, the Laplacian denoising model
To observe the feature extraction effect of the network, heatmaps are adopted to vis-
and the CBAM can reduce speckle noise and enhance high-frequency information and
ualize the effect of the network. As shown in Figure 7, the Laplacian denoising model and
effective features. The modules improve the detection effect, especially for the small, dense,
the CBAM
parallel, andcan reducetargets.
inshore speckle Innoise androw
the first enhance high-frequency
of Figure 7, the Laplacianinformation
denoising andmodeleffec-
tive features. The modules improve the detection effect, especially for the
effectively highlights the small targets in the image with speckle noise through denoising small, dense,
parallel, and inshore
and improves targets.
the feature In the effect
extraction first row of Figure
of small 7, In
targets. thethe
Laplacian denoising
second and model
third rows,
effectively highlights the small targets in the image with speckle noise
the parallel ships in the complex background are separated through the network. The through denoising
and improves
locations of thetheships
feature extractionmore
are predicted effectaccurately.
of small targets. In the
Moreover, as second
shown in andthethird rows,
fourth
the
row,parallel shipsextracts
the LPDNet in the complex background
the features are separated
of the densely through
arranged inshore theand
ships network. The lo-
pays more
cations
attention of to
thethem.
shipsAtare predicted
the same time, more
the accurately. Moreover,
false alarm rate of small,assuspected
shown intargets
the fourth row,
in the
figure
the is reduced.
LPDNet extracts the features of the densely arranged inshore ships and pays more
attention to them. At the same time, the false alarm rate of small, suspected targets in the
3.4.2. Analysis of Detection Results
figure is reduced.
As shown in Figure 8, the images contain different scenes and resolutions. The
multi-level Laplacian pyramid denoising model can remove the speckle noise with large
amplitude, highlight the small targets, and reduce missed detection. In the first and
second rows of Figure 8, LP Yolox-tiny improves the accuracy of small target detection
in the case of high noise. The denoising part is helpful in detecting small targets in noisy
scenes as the confusing noise is depressed. At the same time, high-frequency information
in the three-channel image provides more boundary information that can improve the
effect of target detection. As shown in the third row, high-frequency contour features are
beneficial to the extraction of target and coast features, which improve the detection effect of
inshore targets. The CBAM can adaptively refine features, focus on important features, and
suppress unnecessary features by deducing the attention map along the channel and space
dimensions to increase the representation ability. As shown in the fourth and fifth rows,
Sensors 2023, 23, 6084 10 of 16

Sensors 2023, 23, x FOR PEER REVIEW 10 of 16

the CBAM makes the network pay more attention to the target features. The detection of
densely arranged inshore ships and parked ships in the port is more accurate. The detection
results in Table 2 further demonstrate the effectiveness of each module.

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

Figure 7.
Figure 7. Some
Some visualization
visualization results.
results. (a)
(a) Ground
Ground truth;
truth; (b)
(b) Yolox-tiny;
Yolox-tiny; (c)
(c) LP
LP Yolox-tiny;
Yolox-tiny; (d)
(d) LPDNet.
LPDNet.

3.4.2. 2.
Table Analysis of Detection
The results of ablationResults
experiments.

Methods Recall Precision AP F1 Params FPS

Yolox-tiny 86.81% 92.94% 93.68% 0.9 5,032,866 30.94
LP Yolox-tiny 91.03% 94.67% 96.25% 0.93 5,144,200 25.19
LPDNet (proposed) 90.66% 95.74% 97.14% 0.93 5,150,158 24.68
 Sensors
Sensors 2023,
2023, 23,23, x FOR PEER REVIEW
6084 1111ofof16
16

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

Figure8.8. The
Figure The results
results of
of ablation
ablationexperiments.
experiments.(a)(a)
Ground truth;
Ground (b) (b)
truth; Yolox-tiny; (c) LP
Yolox-tiny; (c)Yolox-tiny; (d)
LP Yolox-tiny;
LPDNet.
(d) LPDNet.
 Sensors 2023, 23, 6084 12 of 16

3.5. Comparison Experiment

In this section, the comparison experiments are carried out on the SSDD and AIR
SARShip-1.0, including SSD, CenterNet, Yolo v4-tiny, Yolo v7, and Yolox-tiny. Since the
two datasets contain images of various scenes, polarization modes, and resolutions, the
experimental results are robust and universal. The comparison experiments are carried out
on the same train and test datasets, and the default parameters of the models are used.
The results of the experiments are shown in Tables 3 and 4. On the SSDD, the AP of the
LPDNet achieves 97.14%, which is 15.13%, 5.91%, 6.92%, 7.34%, and 3.46% higher than the
SSD, CenterNet, Yolo v7, Yolo v4-tiny, and Yolox-tiny, respectively. The parameters of SSD,
CenterNet, and Yolo v7 are several times the proposed algorithm, while the lightweight
methods, Yolo v4-tiny and Yolox-tiny, have similar numbers of parameters as the proposed
algorithm. Meanwhile, lightweight algorithms take less time to test than the algorithm
proposed, but the accuracies of the algorithms are inferior. On the AIR SARShip-1.0 dataset,
the comparison algorithms have similar effects. The methods’ precision-recall (PR) curves
on the SSDD and AIR SARShip-1.0 are shown in Figures 9 and 10, respectively. The
results of the LPDNet are excellent. On the one hand, the intelligent denoising module
adaptively denoises each image to provide a clear image for the feature extraction module.
At the same time, high-frequency images provide more explicit boundary information
for the feature extraction process and strengthen the feature extraction effect. On the
other hand, CBAM highlights the effective features, suppresses the invalid features, and
enhances the target detection effect. The LPDNet improves both the precision and recall
considerably, reaching the best performance in the mentioned algorithms, while considering
Sensors 2023, 23, x FOR PEER REVIEW
both lightweight and detection speed. To sum up, the algorithm in this article further
improves the accuracy of target detection on the premise of ensuring lightweight, which
indicates a better detection effect.

Figure 9. Precision-recall (PR) curves of different algorithms on the SSDD.

Figure 9. Precision-recall (PR) curves of different algorithms on the SSDD.
Sensors 2023, 23, 6084 13 of 16

Table 3. The results of different methods on SSDD.

Methods Recall Precision AP F1 Params FPS

SSD 56.41% 95.36% 82.01% 0.71 23,612,246 8.161
CenterNet 72.98% 97.49% 91.23% 0.83 32,664,197 18.93
Yolo v7 77.29% 90.75% 90.22% 0.83 37,194,710 14.88
Yolo v4-tiny 83.86% 89.76% 89.8% 0.87 5,874,116 54.05
Yolox-tiny 86.81% 92.94% 93.68% 0.9 5,032,866 30.94
LPDNet (proposed) 90.66% 95.74% 97.14% 0.93 5,150,158 24.68

Table 4. The results of different methods on AIR SARShip-1.0.

Methods Recall Precision AP F1 Params FPS

SSD 76.73% 82.2% 86.11% 0.79 23,612,246 7.71
CenterNet 67.07% 94.02% 90% 0.78 32,664,197 14.17
Yolo v7 68.64% 84.06% 81.61% 0.76 37,194,710 14.65
Yolo v4-tiny 61.77% 77.7% 72.98% 0.69 5,874,116 47.85
Yolox-tiny 86.15% 83.83% 89.96% 0.85 5,032,866 29.94
Figure
LPDNet9.(proposed)
Precision-recall (PR)
88.64% curves of different
84.88% algorithms
92.19% 0.87 on the SSDD.26.76
5,150,158

Figure 10. Precision-recall (PR) curves of different algorithms on the AIR SARShip-1.0.
Figure 10. Precision-recall (PR) curves of different algorithms on the AIR SARShip-1.0.
3.6. Migration Ability
To test the migration
3.6. Migration Ability ability of LPDNet, we performed an experiment on the HRSID.
We divided the entire dataset according to the dataset division of the original paper. Table 5
showsTo thetest the migration
detection results on ability of LPDNet,
the HRSID. we method
The proposed performed an experiment
can achieve state-of- on t
We divided the entire dataset according to the dataset division of the 93%,
the-art results, with AP, F1, recall, and precision values of 93.09%, 0.89, 85.99%, and original pa
respectively. The results show that the proposed method has a strong migration ability.
5 shows the detection results on the HRSID. The proposed method can achiev
the-art results, with AP, F1, recall, and precision values of 93.09%, 0.89, 85.99%
respectively. The results show that the proposed method has a strong migratio

Table 5. The results of different methods on HRSID.

Method Recall Precision AP

Faster R-CNN 77.5% 88.8% 78.2%
Sensors 2023, 23, 6084 14 of 16

Table 5. The results of different methods on HRSID.

Method Recall Precision AP F1

Faster R-CNN 77.5% 88.8% 78.2% 0.83
Cascade R-CNN 79.3% 89.9% 79.2% 0.84
SSD 85.3% 87.4% 88.8% 0.86
CenterNet 87.4% 81.8% 86.3% 0.85
RetinaNet 83.8% 69.8% 82.5% 0.76
FCOS 79.5% 91.9% 86.6% 0.85
Yolox-tiny 82.33% 94.02% 91.08% 0.88
LPDNet (proposed) 85.99% 93% 93.09% 0.89

4. Conclusions
In this article, a new lightweight SAR ship detector, LPDNet, is proposed. Based on
the multiresolution characteristics of Laplacian pyramid transformation, a CNN-based
method is used to obtain the denoising threshold from the global characteristics of each
image in a supervised way to achieve intelligent denoising. At the same time, channel
modeling is adopted to make up for the loss of information during denoising, enhance
the contour features, and improve the detection effect. Moreover, CBAM is introduced
into the feature fusion module to enhance the ability of feature representation and further
improve the network performance. Experiments show that the proposed network achieves
a balance between lightweight and detection accuracy and has an excellent detection effect.
In the future, the generalization performance of LPDNet will be studied, including
validation on large scene datasets, and the effects of image resolution, polarization mode,
and other conditions on target detection. We will lighten the detector without sacrificing
accuracy and obtain better detection results. Further, we will explore the application of the
multi-level Laplacian pyramid denoising model in target recognition, segmentation, and
other fields. In addition, more intelligent denoising methods will be explored. We will focus
on combining traditional denoising methods and CNNs so that the traditional denoising
methods can develop intelligence and embeddability to serve in image interpretation.

Author Contributions: Methodology, C.Z. and J.D.; Data curation, C.Z.; Software, C.Z., H.C. and
C.F.; Writing original draft, C.Z. and J.D.; Validation, X.F. All authors have read and agreed to the
published version of the manuscript.
Funding: This work was supported by the 111 Project of China under Grant B14010.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Yang, R.; Hu, Z.; Liu, Y.; Xu, Z. A Novel Polarimetric SAR Classification Method Integrating Pixel-Based and Patch-Based
Classification. IEEE Geosci. Remote Sens. Lett. 2020, 17, 431–435. [CrossRef]
2. Qin, R.; Fu, X.; Lang, P. PolSAR Image Classification Based on Low-Frequency and Contour Subbands-Driven Polarimetric SENet.
IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 4760–4773. [CrossRef]
3. Brusch, S.; Lehner, S.; Fritz, T.; Soccorsi, M.; Soloviev, A.; van Schie, B. Ship Surveillance with TerraSAR-X. IEEE Trans. Geosci.
Remote Sens. 2011, 49, 1092–1103. [CrossRef]
4. Wei, X.; Wang, X.; Chong, J. Local region power spectrum-based unfocused ship detection method in synthetic aperture radar
images. J. Appl. Remote Sens. 2018, 12, 016026. [CrossRef]
5. Huo, W.; Huang, Y.; Pei, J.; Zhang, Q.; Gu, Q.; Yang, J. Ship detection from ocean SAR image based on local contrast variance
weighted information entropy. Sensors 2018, 18, 1196. [CrossRef]
6. Yang, M.; Guo, C. Ship Detection in SAR Images Based on Lognormal ρ-metric. IEEE Geosci. Remote Sens. Lett. 2018, 15, 1372–1376.
[CrossRef]
Sensors 2023, 23, 6084 15 of 16

7. Eldhuset, K. An automatic ship and ship wake detection system for spaceborne SAR images in coastal regions. IEEE Trans. Geosci.
Remote Sens. 1996, 34, 1010–1019. [CrossRef]
8. Shi, H.; Zhang, Q.; Bian, M.; Wang, H.; Wang, Z.; Chen, L.; Yang, J. A novel ship detection method based on gradient and integral
feature for single-polarization synthetic aperture radar imagery. Sensors 2018, 18, 563. [CrossRef]
9. Schwegmann, C.P.; Kleynhans, W.; Salmon, B.P. Manifold adaptation for constant false alarm rate ship detection in South African
oceans. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2015, 8, 3329–3337. [CrossRef]
10. Pappas, O.; Achim, A.; Bull, D. Superpixel-Level CFAR Detectors for Ship Detection in SAR Imagery. IEEE Geosci. Remote Sens.
Lett. 2018, 15, 1397–1401. [CrossRef]
11. Li, T.; Liu, Z.; Xie, R.; Ran, L. An Improved Superpixel-Level CFAR Detection Method for Ship Targets in High-Resolution SAR
Images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 184–194. [CrossRef]
12. Ao, W.; Xu, F.; Li, Y.; Wang, H. Detection and Discrimination of Ship Targets in Complex Background from Spaceborne ALOS-2
SAR Images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 536–550. [CrossRef]
13. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You Only Look Once: Unified, Real-Time Object Detection. In Proceedings of the
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 779–788.
14. Liu, W.; Anguelov, D.; Erhan, D.; Szegedy, C.; Reed, S.; Fu, C.Y.; Berg, A.C. SSD: Single shot multibox detector. In Proceedings of
the European Conference on Computer Vision (ECCV), Amsterdam, The Netherlands, 11–14 October 2016; pp. 21–37.
15. Lin, T.Y.; Goyal, P.; Girshick, R.; He, K.; Dollár, P. Focal Loss for Dense Object Detection. IEEE Trans. Pattern Anal. Mach. Intell.
2020, 42, 318–327. [CrossRef]
16. Tian, Z.; Shen, C.; Chen, H.; He, T. FCOS: Fully Convolutional One-Stage Object Detection. In Proceedings of the 2019 IEEE/CVF
International Conference on Computer Vision, Seoul, Republic of Korea, 27–28 October 2019. [CrossRef]
17. Girshick, R.; Donahue, J.; Darrell, T.; Malik, J. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, USA, 23–28 June 2014;
pp. 580–587. [CrossRef]
18. Ren, S.; He, K.; Girshick, R.; Sun, J. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. IEEE
Trans. Pattern Anal. Mach. Intell. 2017, 39, 1137–1149 . [CrossRef] [PubMed]
19. He, K.; Gkioxari, G.; Dollár, P.; Girshick, R. Mask R-CNN. IEEE Trans. Pattern Anal. Mach. Intell. 2020, 42, 386–397. [CrossRef]
20. Cai, Z.; Vasconcelos, N. Cascade R-CNN: Delving into High Quality Object Detection. In Proceedings of the 2018 IEEE/CVF
Conference on Computer Vision and Pattern Recognition, Salt Lake City, UT, USA, 18–23 June 2018; pp. 6154–6162. [CrossRef]
21. Li, J.; Qu, C.; Shao, J. Ship detection in SAR images based on an improved faster R-CNN. In Proceedings of the 2017 SAR in Big
Data Era: Models, Methods and Applications (BIGSARDATA), Beijing, China, 13–14 November 2017; pp. 1–6. [CrossRef]
22. Wei, S.; Zeng, X.; Qu, Q.; Wang, M.; Su, H.; Shi, J. HRSID: A high-resolution SAR images dataset for ship detection and instance
segmentation. IEEE Access 2020, 8, 120234–120254. [CrossRef]
23. Huang, L.; Liu, B.; Li, B.; Guo, W.; Yu, W.; Zhang, Z.; Yu, W. OpenSARShip: A Dataset Dedicated to Sentinel-1 Ship Interpretation.
IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2018, 11, 195–208. [CrossRef]
24. Li, Y.; Zhang, S.; Wang, W.-Q. A Lightweight Faster R-CNN for Ship Detection in SAR Images. IEEE Geosci. Remote Sens. Lett.
2022, 19, 4006105. [CrossRef]
25. Cui, Z.; Li, Q.; Cao, Z.; Liu, N. Dense Attention Pyramid Networks for Multi-Scale Ship Detection in SAR Images. IEEE Trans.
Geosci. Remote Sens. 2019, 57, 8983–8997. [CrossRef]
26. Woo, S.; Park, J.; Lee, J.Y.; Kweon, I.S. CBAM: Convolutional block attention module. In Proceedings of the European Conference
on Computer Vision, Munich, Germany, 8–14 September 2018. [CrossRef]
27. Wan, H.Y.; Chen, J.; Huang, Z.X.; Xia, R.F.; Wu, B.C.; Sun, L.; Yao, B.D.; Liu, X.P.; Xing, M.D. AFSar: An Anchor-Free SAR
Target Detection Algorithm Based on Multiscale Enhancement Representation Learning. IEEE Trans. Geosci. Remote Sens. 2022,
60, 5219514. [CrossRef]
28. Ge, Z.; Liu, S.; Wang, F.; Li, Z.; Sun, J. YOLOX: Exceeding YOLO Series in 2021. arXiv 2021, arXiv:2107.08430v2.
29. Qi, X.; Lang, P.; Fu, X.; Qin, R.; Dong, J.; Liu, C. A Regional Attention-based Detector for SAR Ship Detection. Remote Sens. Lett.
2022, 13, 55–64. [CrossRef]
30. Shiqi, C.; Ronghui, Z.; Jun, Z. Regional Attention-based Single Shot Detector for SAR Ship Detection. J. Eng. 2019, 2019, 7381–7384.
[CrossRef]
31. Mao, Y.; Yang, Y.; Ma, Z.; Li, M.; Su, H.; Zhang, J. Efficient low-cost ship detection for SAR imagery based on simplified U-Net.
IEEE Access 2020, 8, 69742–69753. [CrossRef]
32. Ronneberger, O.; Fischer, P.; Brox, T. U-Net: Convolutional networks for biomedical image segmentation. arXiv 2015,
arXiv:1505.04597.
33. Liu, L.; Jia, Z.; Yang, J.; Kasabov, N.K. SAR Image Change Detection Based on Mathematical Morphology and the K-Means
Clustering Algorithm. IEEE Access 2019, 7, 43970–43978. [CrossRef]
34. Farhadiani, R.; Homayouni, S.; Safari, A. Hybrid SAR Speckle Reduction Using Complex Wavelet Shrinkage and Non-Local
PCA-Based Filtering. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2019, 12, 1489–1496. [CrossRef]
35. Zhang, J.; Li, W.; Li, Y. SAR Image Despeckling Using Multiconnection Network Incorporating Wavelet Features. IEEE Geosci.
Remote Sens. Lett. 2020, 17, 1363–1367. [CrossRef]
Sensors 2023, 23, 6084 16 of 16

36. Liu, S.; Lei, Y.; Zhang, L.; Li, B.; Hu, W.; Zhang, Y.-D. MRDDANet: A Multiscale Residual Dense Dual Attention Network for SAR
Image Denoising. IEEE Trans. Geosci. Remote Sens. 2022, 60, 5214213. [CrossRef]
37. Sun, X.; Wang, Z.; Sun, Y.; Diao, W.; Zhang, Y.; Fu, K. AIR-SARShip-1.0: High-resolution SAR ship detection dataset. J. Radars
2019, 8, 852–862. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.