2019 IEEE 4th International Conference on Cloud Computing and Big Data Analytics

A Ship Detection Method based on Recurrent Neural Network in a Marine Radar

System

Jie Z Yufen S
Computer Science and Technology department Computer Science and Technology department
Wuhan University of Technology Wuhan University of Technology
Wuhan, China Wuhan, China
e-mail: [email protected] e-mail: [email protected]

Abstract—A Ship Detection Method based on Recurrent recognition method based on feedforward neural network.
Neural Network in a Marine Radar System is investigated. However, the author has not deleted the feature dimension
Considering that the characteristics of the single-frame radar which exists the problem, and the performance of the model
image are not obvious, we selected image sequences in other classifiers needs to be further expanded. Some
concatenated by single-frame radar images as research objects studies have also selected image sequences that are
and designed a coding mode to complete the mapping of radar concatenated by single-frame radar images as research
targets to vectors. Our model based on RNN made a good objects. Weigen Huang [3] proposed a ship target
performance on the dataset made by ourselves. recognition method based on Kalman filter in 2014 and
realized path prediction through similarity calculation of
Keywords-radar image; sequence; RNN; detection Normalized Cross-correlation function. The focus of this
I. INTRODUCTION model is on the identification and path tracking of a single
ship target in a continuous frame radar image, without
The shipborne radar target recognition, that is, radar effectively distinguishing ship targets and clutter.
detection and identification of the target against a Considering that the characteristics of the single-frame
background of clutter and interference, is of great radar image are not obvious, this paper intends to use the
significance in maritime surveillance, ship route planning, target sequence composed of continuous N-frame radar
risk warning and so on. At this stage, the target recognition image as the feature extraction object. Through RNN, we
of shipborne radar still has great difficulty. On the one hand, hoped that the context information of the radar image target
the radar image is wholly composed of the strength of the sequence could be excavated to enhance the credibility of
target echo, and the number of its features is far less than the radar target recognition, and effectively identify the ship
that of the optical image. The existence of factors such as target and clutter.
the dense distribution of sea targets and the dynamic
variation of sea clutter has increased the difficulty of target II. RELATED WORK
recognition of shipborne radar. On the other hand, shipborne
The research on target recognition of shipborne radar
radar is susceptible to weather and various clutter, noise and
has been continuously developed since the 1950s as a
co-channel interference in practical applications. Target
specific application of radar target recognition. Over the
echo signals and radar images are severely affected,
past decades, with the continuous development of radar
resulting in the inability to distinguish ship targets and
technology and continuous enrichment and improvement of
clutter
At this stage, the field of shipborne radar target various basic theories, shipborne radar target recognition
recognition mainly uses single-frame radar signals as has achieved excellent results. In recent years, the rapid
research objects, whether it is a one-dimensional range growth of techniques such as machine learning and deep
image, SAR image or other forms of radar echo signals. Yu learning has also provided a new solution for shipborne
Zhi-fu [1] proposed a fusion recognition method based on radar target recognition.
PSO-BP neural network in 2012. The BP neural network In 1983, PE Zwicke [4] developed a digitally
optimized by particle swarm optimization (PSO) algorithm implemented Mellin transform and applied it to radar target
was applied to radar image recognition. However, the recognition. The translation and scale invariance of the
problem with the model is that the feature parameters with Mellin transform provides a way to extract features from the
discrete values are not selected, so the recognition efficiency radar target profile that are insensitive to the radar's vertical
of the negative samples needs to be further improved. A and horizontal angles. The authors use the direct extension
problem with the model is that the feature parameters with of the Mellin basic definition to implement the Mellin
discrete values are not selected, so the recognition of the transform. The new method eliminates several problems
negative samples needs to be further improved. In 2013, such as the explosion of the sample number caused by
Nedyalko Petrov [2] designed an encoded mode for radar exponential sampling and the error mixed in the
echo signal eigenvectors and proposed a radar target transformation caused by the calculation of the interpolation

978-1-7281-1410-1/19/$31.00 ©2019 IEEE 548

Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:36:30 UTC from IEEE Xplore. Restrictions apply.
value and the correction term in the Mellin transform based All the data of the RNN_Radar model training and
on the fast Fourier transform. The Mellin transform applied testing are from the on-site collection of the shipborne 4G
to the radar profile of the ship achieved good recognition radar on the banks of the Yangtze River. After analyzing
and eliminated the low-pass filtering characteristics found in and comparing various situations of clutter and ship target
other models. appearing in radar image, we summarize the characteristics
In 2011, Z Yin et al. [5] proposed a radar target of clutter contour in radar image, and then designed an
recognition model based on a single parameter dynamic encode mode to complete the mapping of radar targets to
search (SPDS) algorithm. The SPDS algorithm is a modified vectors, which will be introduced in detail in the part of the
algorithm of the BP network that allows only one of all pre-processing network. The characteristics of the clutter
parameters in the network to be changed during each period contour are as follows:
of the search step to ensure accurate one-dimensional search 1. The contour area of the clutter in the radar is always
is performed, which can overcome the local minimum and at a small value compared to the ship's target.
long training time of the BP neural network. Compared with 2. Compared to the first occurrence of a ship's target
the test results of the K-Nearest Neighbor classifier, the which must be located at the edge of the radar image,
author demonstrates that the SPDS algorithm can more the location where the clutter first appears is random.
accurately identify the radar target, and the recognition rate 3. Clutter often does not appear alone, it is easier to
is close to 100% under sufficient training time and simple find clutter contours near the ship's target, or due to
data conditions. the frequency refresh during radar acquisition, the
Wang C, Zhang H, Wu F et al. [6] used the COSMO- clutter will appear as a strip from the radar center as
SkyMed strip image of 3M resolution in HIMAGE model to a starting point.
4. Compared with the ship target which can maintain a
study the characteristics of merchant ships and proposed a
relatively stable state in the continuous frame radar
ship target classification method. After pre-processing, the
image, the state of the clutter in the radar image is
feature vector is estimated by the mean of the kernel density
unstable, that is, several frames appear briefly in the
estimates, the three structural features, and the average
radar sequence and then disappear.
backscatter coefficient. The author used support vector 5. The positional state of the clutter generally does not
machine (SVM) for the classification of feature vectors of change in its series.
ship targets and compared the results with the operation of
K-nearest neighbour algorithm (K-NN) and minimum
distance classifier (MDC). Preliminary results show that the
combination of the average of the nuclear density estimates
and the average backscattering coefficient has a good
classification of three types of ships.
In many cases, the models of machine learning
combined with traditional algorithms such as fuzzy systems,
clustering algorithms, wavelet packets, and Kalman filters
are more accurate and efficient in ship targets recognition [7,
8, 9].
III. THE RNN_RADAR ARCHITECTURE
Our model is inspired by the RNN Language Model [10] Figure 1. The RNN_radar Architecture
proposed by Mikolov in 2011. We will compare the single-
frame radar image which is difficult to extract effective IV. THE PRE-PROCESSING NETWORK
features to the words in RNNLM, and hope to explore the The pre-processing network is mainly used for pre-
useful information in the radar image sequence through processing of radar data in this model. The target contour in
RNN training, so as to effectively distinguish the clutter and the radar image is mapped into the feature vector
the ship target in the radar image. The network structure of representing the contour as the input of the RNN network,
the RNN_Radar model proposed in this paper is divided which promotes the model faster and more accurate. The
into two parts: A pre-processing network for radar target pre-processing network includes steps such as radar target
feature extraction and an RNN network for model data contour extraction, feature vector extraction, and contour
training and testing. The pre-processing network includes sequence extraction. The input to the pre-processing
steps of radar target contour extraction, feature vector network is a continuous-period N-frame radar image, and
extraction, contour sequence extraction, and the like; The the output of the network is a feature vector representing
training and test network is mainly based on RNN, and the each target contour and a time-ordered contour sequence for
corresponding radar images are classified according to the different ship targets. The complete pre-processing network
model training results. The complete network architecture is architecture is shown in Fig. 2. Next, I will detail the
shown in Fig. 1. various parts of the pre-processing network.

549
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:36:30 UTC from IEEE Xplore. Restrictions apply.
 feature vector. We borrowed the idea that the RNN model
used dimensional reduction for multidimensional input in
NLP and selected the vector coding method combining
Label Encoding and binary coding to extract 15D for each
input target contour. The feature vector is used for the
calculation of subsequent models.
The first ten-dimensional features of each set of feature
vectors are tagged. The content is the HU moment of the
target contour (HU [0]-HU [6]), the centroid of the target
contour, and the area, representing the physical meaning of
the target contour. The first ten-dimensional features are
used to calculate the similarity between the target contours
Figure 2. The Pre-processing Network in the process of contour sequence extraction. The last five-
dimensional features of feature vectors are binary coding.
A. Target Contour Extraction According to the five characteristics summarized above,
The pre-processing network reads the continuous N- five thresholds are set after several rounds of testing. Each
frame radar image from the corresponding folder. After feature is set 0 or 1 according to whether it reaches the
completing the grey-scale conversion, we use OpenCV’s threshold. The specific meanings of the five-dimensional
findContours () function to extract the contour of each frame features are: 1.Whether the area of the target contour is
of the radar image and complete the preliminary mapping smaller than the limited area S; 2.Whether the target contour
from the target contour to the contour vector. After the is at the edge of the image; 3.Whether there are other
garbage outline with a contour area of 0 is removed, the contours within the range of the limited distance d below the
remaining contour vectors are input to the next stage for target contour; 4.Whether the length of the sequence in
feature vector extraction. which the target contour is located is smaller than the
limited length L; 5.Whether the displacement of the target
B. Feature Vector Extraction contour and the previous one is smaller than limited
According to the different characteristics of the above- displacement X. Table I gives a sample of feature vector
mentioned summarized clutter and ship target in the radar coding.
image, we design the composition rules of the target contour

TABLE I. A SAMPLE OF FEATURE VECTOR CODING

ID Class HU Centroid Area Area smallerLength Displacement At the edge? Another

than S? less than L? less than X? contour?

257 Ship ….. (10.1, 1.5) 25 0 0 0 1 0

632 Clutter ….. (255.0,97.5) 5 1 1 1 0 1

955 Clutter ….. (98.6,780.2) 13 0 1 1 0 1

C. XML File Reading D. Contour Sequence Extraction

In training mode, the extra step of the preprocessing The pre-processing network extracts all the contour
network is to read the XML file of the training dataset. feature vectors into the contour sequence of its
Since the output of the RNN is the category of the target corresponding target by contour sequence extraction, and
contour, the category information read by the XML file will finally obtains complete contour sequences of different
be used for the calculation of the RNN backpropagation. In targets (ships or clutter) for the input of the train or test in
the process of making the training dataset, we used the the RNN. The extraction method of the contour sequence
LabelImg annotation tool to mark the position and the uses the cosine vector similarity calculation between the two
category information of the target contour for each radar target contour feature vectors. For the contour of a target at
image used for training, and save it as the corresponding the current moment, the similarity of all the contours in the
XML file. Therefore, we use the library TinyXML to read radar image of the next frame is calculated, and the contour
the XML file corresponding to each frame radar image and with the highest similarity which reaches a certain threshold
find the matching target contour in the data structure is selected and added to the contour sequence of the object.
according to its location, add the category information to its If the contour with the highest similarity does not reach the
structure. threshold, the sequence terminator is activated, and the
contour sequence extraction of the object is terminated.

550
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:36:30 UTC from IEEE Xplore. Restrictions apply.
 V. RNN
Because the characteristics of single-frame radar image
are not obvious and it is difficult to distinguish from clutter,
this model uses RNN as the classifier. We hoped that the
radar target recognition credibility could be enhanced by
extracting the context information of the radar target
contour sequence to achieve an effective distinction between
the ship target and clutter. In this model, the input of the
RNN is a target contour feature vector of 15-dimensional,
and the output is a 1-dimensional size target contour
category, where 0 represents the target contour as the ship
Figure 4. The RNN structure at a certain moment
target, and 1 represents the clutter. Since the input vector
dimension of the RNN is relatively low and there is no
The loss function of backpropagation in the model uses
situation where the dimensional explosion affects the
the standard logistic regression loss function, and the update
performance of the model, we use the basic RNN unit
of the model parameters such as weights in the model is
structure for training.
realized by the gradient descent. To prevent the model from
the gradient explosion in the back-propagation due to
abnormal data, we adopt the method of gradient clipping to
suppress. When the amplitude of the gradient is greater than
the set threshold, the gradient will be equal to the threshold,
that is, the gradient stops increasing.
B. Category Annotation
In the test mode, the model will classify the
corresponding target contour according to the output of the
RNN, which complete the final step of the end-to-end
output of the entire model. According to the results of
several rounds of model training tests, the boundary value
Figure 3. The RNN structure between the ship target and the clutter is stable at around
0.6590, so we set the threshold to 0.6590. For the contour
Fig. 3 shows the RNN structure under a complete with the value of the output layer higher than 0.6590, we use
sequence of a target, where w(t), s(t), s(t-1), y(t) putText () in the library OpenCV to tag it as clutter, and the
respectively represents the input layer, the hidden layer of contour whose output value is less than or equal to 0.6590 is
the current and the previous moment, the output layer, and tagged as a ship.
U, V, W are the corresponding weight matrixes. In the
actual model, w(t) and s(t-1) are combined into the input VI. EXPERIMENT
layer.
A. Data Preparation
A. Forward Propagation and Back Propagation In the field of radar image target recognition, there is no
Like other RNN models, the RNN of our model includes more general training data set for the academic research of
two parts: forward propagation and backpropagation. The shipborne radar target recognition. Considering the
forward propagation is mainly used for the identification influence of different data sets on model parameters and
between ship targets and clutter. Also, the information recognition effects, our team came to the Yangtze River
contained in the current feature vector is passed to the next. with 4G radar. Field data collection was carried out on the
Backpropagation is mainly used to update the parameters edge of Wuchang River Beach. To eliminate the influence
and weights of the model. of some factors in the acquisition process on the
According to the structure diagram of the RNN at a experimental results, we removed some radar images with
certain moment given in Fig. 4, we can clearly understand radar noise and fusion between two targets in the dataset.
how we calculate the category probability for the radar The 1000 consecutive radar images with clear radar
target contour at each moment. Considering that the contours and less radar noise are selected for model training
problem dealt with this model is a two-class problem (the and testing. At the same time, in order to avoid the
distinction between ship target and clutter), the target output interference of the complex radar contours on the bridge
function uses the sigmoid function, so the RNN output pier and both sides of the Yangtze River, we selected the
formula at each moment is: target area of the same size that contains only the ship and
the buoy (i.e. non-hull) and uniformly cut the radar image.
s( ) () s - (1) For the data used for model training, we use the labelImg
() s (2) tool to mark the location and the information of category of
each target in the radar image.
() (3)

551
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:36:30 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. THE TEST RESULTS

Total Effective Accuracy Precision Ship Clutter

contours contours

2271 2053 0.81976 0.91349 0.86671 0.55821

B. Training and Result

Since the training of RNN does not have high
requirements for the performance of the computer, all the
training of the model was done on the NVidia 1050Ti GPU.
We used a 753-frame continuous radar image, extracted a Figure 6. Test result of the radar image with strip clutter
total of 9742 contours, and composed 669 target contour
sequences to complete the training in the model. The 2271
contours of the 262-frame continuous radar image acquired
on a different day were tested on the trained model.
Table II lists the indicators of the test results, in which
the recognition rate of the ship target reached 86.671%, and
the recognition rate of the clutter reached 55.821%.
Some of the representative radar images in the test
results are listed below. Fig. 5 and Fig. 6 show that some
radar images with visible negative samples have good
discrimination between ship targets and clutter. Fig. 7 and
Fig. 8 show some issues in the test such as misdetections,
missed inspection and repeated detection of targets. Figure 7. Misdetection of the target contour (clutter at the bottom
should be the ship target)
VII. CONCLUSION
The model proposed in this paper designs the
corresponding feature vector according to the different
characteristics of clutter and ship target in radar image and
uses the recurrent neural network RNN to process the
sequence data of radar images and identify the clutter and
ship target. The model made good performance on the
datasets obtained from field surveys, but there are still some
limitations. Firstly, there are few authoritative datasets of
one-dimensional radar image, and the performance of this
model on other datasets needs to be further expanded;
Secondly, our dataset eliminates the influence of the
complex contours of the Yangtze River bridge pier and Figure 8. Missing detection of target contour (A target contour is not
cross-strait, so our model's performance in the case of labeled)
increasing the number of categories needs to be further
expanded. ACKNOWLEDGEMENT
This research is supported by the open fund from
National Engineering Research Center for Water Transport
Safety.
REFERENCES
[1] Yu Z F, Li J W, Liu K. Radar Emitter Recognition Based on PSO-BP
Network[J]. Aasri Procedia, 2012, 1(4):213-219.
[2] Petrov N, Jordanov I, Roe J. Radar Emitter Signals Recognition and
Classification with Feedforward Networks[J]. Procedia Computer
Science, 2013, 22:1192-1200.
[3] Huang W, Liu B. A ship detection and tracking method with time
sequential shipborne radar imagery[J]. Proceedings of SPIE - The
Figure 5. Radar image with strip clutter International Society for Optical Engineering, 2009, 7495:71050Q-
71050Q-7.

552
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:36:30 UTC from IEEE Xplore. Restrictions apply.
[4] Zwicke P E, Kiss I. A new implementation of the mellin transform [7] Granger E, Rubin M, Grossberg S, et al. A what-and-where fusion
and its application to radar classification of ships[J]. IEEE neural network for recognition and tracking of multiple radar
Transactions on Pattern Analysis & Machine Intelligence, 1983, emitters[J]. Neural Netw, 2001, 14(3):325-344.
5(2):191-9. [8] Azimi-Sadjadi M R, Yao D, Huang Q, et al. Underwater target
[5] Yin Z, Yang W, Yang Z, et al. A Study on Radar Emitter Recognition classification using wavelet packets and neural networks[J]. IEEE
Based on SPDS Neural Network[J]. Information Technology Journal, Transactions on Neural Networks, 2000, 11(3):784-794.
2011, 10(4):883-888. [9] Ting C, Wei G, Bing S. A new radar emitter recognition method
[6] Wu F, Wang C, Jiang S, et al. Classification of Vessels in Single-Pol based on pulse sample figure[C]// Eighth International Conference on
COSMO-SkyMed Images Based on Statistical and Structural Fuzzy Systems & Knowledge Discovery. IEEE, 2011.
Features[J]. Remote Sensing, 2015, 7(5):5511-5533. [10] Mikolov T, Chen K, Corrado G, et al. Efficient Estimation of Word
Representations in Vector Space[J]. Computer Science, 2013.

553
Authorized licensed use limited to: Pontificia Universidad Javeriana. Downloaded on October 23,2023 at 03:36:30 UTC from IEEE Xplore. Restrictions apply.