

Feature Extraction and Recognition of Rotational Target under the

Sea Background

Bing Zhu

, Weixin Gao

, Yali Qin

, Wenfeng Li

and Xianglong Kong

School of Electronic Engineering, Xian Shiyou University, No.18 Dianzi er Road, Xian Shannxi, China

No.210 Institute China Aerospace Science and Technology Corporation, No.8 Dianzi yi Road, Xian Shannxi, China

Shanghai Institute of Satellite Engineering, No.3666 Yuanjiang Road, Shanghai, China

Keywords: Sea Clutter Restraint, Time-frequency Transform, Invariant Moment Features, Rotate Plan, Recognition and

Classification.

Abstract: Considering the impact of sea clutter on target classification and recognition, a method based on RBF is

proposed to restrain the actual sea clutter, which can be converted the sea clutter into random noise. After

denosing, a S transform time-frequency approach is used to obtain the two time-frequency distribution images.

They are helicopter and propeller aircraft images with nosie. Then extracted the invariant moment features of

images for target recognition. The simulation results have shown an average accuracy of 85%, which validates

the effectiveness of this method.

1 INTRODUCTION

The signals of sea skimming flying helicopters and

propeller-driven fixed-wing aircrafts are important

types for naval radar to detect and recognize.

Helicopters and propeller aircrafts are equipped with

large long faster rotating rotors. The rotor blades

turning around the rub with periodic high-speed

rotation makes rotor and electromagnetic wave of

radar interact to produce the periodical change of

echo signal in amplitude and phase, which produces

a beneficial feature to identify micro-Doppler

phenomenon. However, sea clutters are serious

constraints on the detectability of target radar echoes

from sea surface or near the surface, so target

identification in the sea conditions is relatively

difficult.

For detection and target recognition of helicopter,

domestic and foreign researchers have carried out

relevant research work. J. Misiurewicz (Misiurewicz

et al., 1997) analyzed various types of helicopters

echo data, founding the rotation effects of the rotor

blades, so the echo data contained scintillation “pulse”

related to rotational speed of the rotor and the number

of rotor blade; G. C. Gaunaard (Gaunaurd and Strifors,

1996) made an effective identification of different

types of targets based on time frequency distribution

by PWVD; Rotander (Rotander and Von Sydow,

1997)proposed to identify the helicopter by the ratio

between the radius rotor of and the number of blade,

however, the analysis is conducted in an environment

which is noise and clutter free; Ding Jianjiang et al.,

analyzed micro-Doppler effects on rotor aircraft,

extracting amplitude, phase, and modulation

characteristics of target echo signal from the time

domain and frequency domain for the classification

and recognition of three types of aircraft.

These studies have not considered target detection

of rotating body in complex conditions. Farina (Gini

and Farina, 1999) detected the rub echo of helicopter

in k distribution clutter background without

considering micro-tremor signal. The RCS of

helicopter rub is generally small, so this method is

only valid for the close-in targets.

In the condition of sea clutter, the spectrum of

target echo signal mixes with sea clutter spectrum,

and the amplitude of target echo signal is not

dominant comparing with amplitude of sea clutter

echo. Using traditional frequency or time domain

processing approach to analyze the target echo signal

interfered by sea clutter is unsatisfactory. Time-

frequency analysis converts radar echo signals from

one-dimensional time or frequency domain to the

joint time-frequency domain for analysis, which can

provide richer target information. The approach of

frequency analysis based on S-transform has the

advantages of Fast Fourier Transform and wavelet

transform, avoiding the disadvantages of both. It has

508

Zhu, B., Gao, W., Qin, Y., Li, W. and Kong, X.

Feature Extraction and Recognition of Rotational Target under the Sea Background.

DOI: 10.5220/0005980805080513

In Proceedings of the 13th International Conference on Informatics in Control, Automation and Robotics (ICINCO 2016) - Volume 1, pages 508-513

ISBN: 978-989-758-198-4

good time-frequency concentration, and does not

contain cross terms (Zhu et al., 2015).

Our proposed framework is shown in figure 1.

Firstly, applicable sea clutter filtering method is

researched; then, combing target recognition method

and spectrogram based on S Transform in the

literature; extracting invariant moments from it, thus

achieving the recognition, extraction and

classification of rotating target under the sea

background.

Figure 1: A flow chart of our method.

Detailed steps are as follows:

(1) Pre-process the original target radar echo data by

filtering and noise reduction;

(2) Suppress preprocessed echo signal with sea clutter,

improving signal-noise ratio;

(3) make s transform for clutter suppressed radar echo

signal, obtaining the image of time-frequency

distribution;

(4) Preprocess the frequency distribution image with

the regularization method;

(5) Extract seven moment invariants features for

regularized image, constituting vector with seven

dimensions;

(6) Feed the moment invariants eigenvector into the

neural network classifier, getting the result by

classification and recognition.

2 FILTERING MODEL OF SEA

CLUTTER BASED ON NEURAL

NETWORK

Traditional description method for sea clutter is to

build classical random statistical model. These

statistical models can only describe changes in sea

clutter from the amplitude; however, the inner

dynamic characteristics generated by sea clutter

cannot be explained. The literature (Haykin and

Puthusserypady, 1997) indicates that quoting chaos

theory into the study of sea clutter is necessary and

feasible. Use phase space reconstruction theory to

obtain internal chaotic dynamics model which

produces sea clutter, then adopting neural network to

study the inherent laws of sea clutter, after that, use

trained neural network to make prediction and

cancellation for sea clutter, and transfer the sea clutter

data with strong amplitude into random noise signal

in strong amplitude, thus achieving the suppression of

sea clutter.

The equation of sea clutter in dynamic systems

can be described as follow:

(1)H(())Xn Xn





(1)

Through the reconstruction of phase space, the

prediction equation of sea clutter can be expressed as:

(1)

(, , )

im i i i m

XFxxx











(2)

Theoretically, if the analytical solution of the

equation is acquired, the predictive value of sea

clutter can be worked out, thus, the suppression of sea

clutter is realized. But solving the analytical solution

is almost impossible. So RBF neural network can be

introduced to predict the equation of sea clutter based

on the known chaotic time series samples of sea

clutter. The schematic diagram of RBF neural

network for prediction is shown in Figure 2. In the

process of forecasting, introduce the observed value

of sea clutter signal at the right side of prediction

equation (2) as RBF input. Input unit is m, which

means that the number of entered layer is equal to the

embedding dimension and RBF network output unit

is appropriate prediction value of sea noise.

Figure 2: Schematic diagram of sea clutter based on RBF

neural network prediction.

Detailed steps of sea clutter suppression based on

neural networks are as follows:



Feature Extraction and Recognition of Rotational Target under the Sea Background

509



(1) Using reconstruction method of phase space to

construct n dimensional training data of, where

the input data is



112

Xxxx





 

，



223 1

Xxxx





 

，…，



nnn mn

Xxx x





 

. Ideal output data:

11m







，

22m







，…，

1nmn







(2) Training the RBF neural network with training

data acquired by the construction of phase space,

and working out network structure after training:

()

YfX

, which



iii mi

Xxx x









，

1imi







(3) Predicting sea clutter by one-step prediction

method, the result can be expressed as

112 1

(, , )

nnnnm

YfXX X





 

(4) Making the predictive value subtract from ideal

value of sea clutter data achieving the cancellation

of sea clutter

1nn







The actual data of sea clutter is used as simulation

data, which length is 1980. Figure 3 shows the

waveform of sea clutter after amplitude normalization.

Figure 3: Sea clutter waveform.

Figure 4: Sea clutter prediction result based on RBF neural

network.

Training and predicting sea clutter based on RBF

neural network. The first 1000 data from collecting

site is used for neural network training, the other 980

is for predicting neural network. With the success of

network training, prediction samples is used for

network prediction of sea clutter, the result is shown

in Figure 4. Network predictionerror, that is, sea

clutter cancellation result is shown in Figure 5.

It can be seen from Figure 5 that RBF neural

network can realize the prediction and cancellation of

sea clutter signal, and transform strong amplitude sea

clutter information into weak amplitude random noise

signal to realize the suppression of sea clutter.

Figure 5: Sea clutter prediction error RBF based on neural

network.

3 TIME-FREQUENCY FEATURE

EXTRACTION AND TARGET

RECOGNITION METHOD

BASED ON S TRANSFORM

Due to the non-stationary, non-Gaussian, and time-

variation of sea clutter signal, traditional time domain

or frequency domain signal processing methods is

difficult to take effect on target detection and

recognition in the condition of sea clutter. The method

of time-frequency analysis converts radar echo

signals from one-dimensional time or frequency

domain to the joint time-frequency domain for

analysis, which can provide richer target information.

The time-frequency distribution of radar echo signals

can be viewed as an image on the plane of two-

dimensional time-frequency distribution, thus the

relevant knowledge of image processing can be

employed to extract target features. For the image, the

moment feature is an extremely valid characteristic,

which mainly characterizes the overall shape of the

target image with invariant features of rotation,

translation, scale and other characteristics. It can

ICINCO 2016 - 13th International Conference on Informatics in Control, Automation and Robotics

510

effectively reflect the essential characteristics of the

image. Time-frequency analysis is used to analyze the

target echo signal interfered by sea clutter, and the

robust features of time-frequency images produced

from various target signal with sea clutter interference

are extracted on the time-frequency plane, which is to

solve the problem of target identification in the sea

condition.

3.1 Time-frequency Analysis of Target

Echo Signal

First of all, time-frequency method is used to analyze

sea clutter. Measured data of sea clutter in the Figure

3 is converted by S Transform, and the corresponding

time-frequency distribution is in the Figure 6. It can

be seen from Figure 6, on the frequency plane, sea

clutter energy mainly concentrates in the low

frequency range, and its energy distribution is

dispersed, with good robustness.

Figure 6: S -Transform frequency spectrums of sea clutter

signals.

3.2 Time-frequency Feature Extraction

and Target Recognition Algorithm

based on S Transform

Moment feature of image usually describes the gene-

ral shape and characteristic, which main idea is to

transform insensitive area- based several moments as

shape characteristics, for the identification of target

image. For a

MN

size digital image, which the

order moment

pq

is defined as:

( , ) , 0,1, 2...

Mijfijpq





(3)

pq

central moment is defined as:

()( )(,)

miijjfij



(4)

where,

10 00 01 00

/, /iM M jM M

(5)

respectively the center of image gray in the horizontal

and vertical directions, also known as the centroid.

Central moment reflects the distribution of gray scale

in the image area comparing with gray center. In order

to get invariant features of moment, the definition of

central moment is normalized as:

/ , ( ) / 2 1, 2,3, 4...

pq pq

mmrpq pq





(6)

Normalized definition of central moment indicates 7

moment invariants with invariability in translation,

scale, and rotation (among which only has translation

and scale invariance). The definition is as (7):

12002

22002 11

33012 2103

43012 2103

5 30 123012 3012 2103

21 03 21 03 30 12 21 03

620023012 210

()4

(3)(3 )

()()

( 3 )( )[( ) 3( ) ]

(3 )( )[3( ) ( ) ]

( )[( ) (

 

 

  

 

  

  

 



 

  

 

    

  

  

3 1130122103

7 21033012 3012 2103

30 12 21 03 30 12 21 03

)] 4 ( )( )

(3 )( )[( ) 3( ) ]

( 3 )( )[3( ) ( ) ]

   

 

   



  

   











(7)

Table 1: Eigenvector of two types of targets in model features library.

Invariant moments of time-frequency distribution based on S transform



Helicopter 5.9741 14.0975 24.0937 22.7050 46.3632 30.2507 46.5574

Feature Extraction and Recognition of Rotational Target under the Sea Background

511



Table 2: Classification results.

Classification rate

Air condition（SNR=8dB） Sea condition（SCR=6dB）

Helicopter 93.33% 90%

Prop 100% 80%

The average recognition rate 96.665% 85%

4 SIMULATION RESULTS

Conduct target echo signal with S Transform

separately from helicopter and propeller aircraft; after

working out its time-frequency distribution, extract

moment invariants feature of image and input them

into neural network classifier for classification and

identification. For simulation, radar and target

parameter are set as same as (Zhu et al., 2015);

according to desired SCR (the ratio of signal to

clutter), adjust data amplitude of sea clutter, and

generate radar echo signal of helicopter and propeller

aircraft in sea condition that SCR=6dB. Collect 45

sets of data for the two objectives respectively and

randomly choose 15 sets of data as training samples,

a total of 30 training samples, another 30 sets of data

as a test sample, a total of 60 test samples, using a

neural network classifier to predict classification,

network prediction results are shown in Fig 7.

Figure 7: Neural network classifiers results.

Table 1 shows eigenvector (average value) of two

types of targets in model library. Figure 2 is based on

S Transform reflecting the recognition result of

moment invariants features in the time-frequency

distribution of target echo signal, which also give the

result of target recognition by using the method of

signal to noise ratio SNR = 8dB (main interference is

Gaussian white noise). The recognition result shows

that it is valid to adopt moment invariant feature to

detect and recognize the target.

5 CONCLUSIONS

Under the sky background, using time-frequency

feature extraction and target recognition method

based on S-transform to recognize target can make

recognition rate of helicopter to 93.33%, the

recognition of propeller aircraft to 100%, and the

average recognition rate of two types of targets to

96.665%.

Under the sea ground, sea clutter interference

makes target recognition rate decline, but after the

introduction of clutter suppression, S-transform based

time-frequency feature extraction and target

recognition method can also reach 90% and 80%

correct recognition rate on helicopters and propeller

aircrafts targets, and average recognition rate of 85%

correspondingly. Time-frequency feature extraction

and target recognition method based on S-transform

are more effective for the detection and recognition of

helicopter and propeller aircraft with a rotating plane,

which are able to reach a certain recognition rate.

ACKNOWLEDGEMENTS

Sponsored by Shaanxi Provincial Science and

Technology Research and Development Program

(2013K09-18), Xi'an shiyou University Youth

Science and Technology Innovation Fund

project(2015BS18), SAST Foundation (Grant No.

SAST2015040).

REFERENCES

Gaunaurd, G. C., & Strifors, H. C. (1996). Signal analysis

by means of time-frequency (Wigner-type)

ICINCO 2016 - 13th International Conference on Informatics in Control, Automation and Robotics

512

distributions-applications to sonar and radar echoes.

Proceedings of the IEEE, 84(9), 1231-1248.

Gini, F., & Farina, A. (1999). Matched subspace CFAR

detection of hovering helicopters. IEEE Transactions

on Aerospace & Electronic Systems, 35(4), 1293-1305.

Misiurewicz, J., Kulpa, K., & Czekala, Z. (1997). Analysis

of recorded helicopter echo. Paper presented at the

Radar 97 (Conf. Publ. No. 449).

Rotander, C. E., & Von Sydow, H. (1997). Classification of

helicopters by the L/N-quotient. 629-633.

Zhu, B, R. X. Yu, & Y. L. Qin. (2015). Analysis and

Simulation of Helicopter Ｒadar Echo Signal Based on

S Transform. Aeronautical Computing Technique, 45(4),

Haykin, S, & Puthusserypady, S., (1997). Chaotic dynamics

of sea clutter. Chaos an Interdisciplinary Journal of

Nonlinear Science, 7(4), 777-802.

Feature Extraction and Recognition of Rotational Target under the Sea Background

513