Garadar target recognition using HRRP in condition of unknown noise

In the paper, the results of thesimulation research of radar target recognition using HRRP in condition of unknown noise where the signal-to-noise ratio (SNR) of training HRRPsand test HRRPs are differentpresented. Athreshold noise redution method and a recognition model are proposed in the paper to improve the recognition performance in condition of unknown noise.

RADAR TARGET RECOGNITION USING HRRP

IN CONDITION OF UNKNOWN NOISE

Nguyen Thanh Hung1,*, Le Hai2, Nguyen Hoang Nguyen2

Abstract: In the paper, the results of thesimulation research of radar target

recognition using HRRP in condition of unknown noise where the signal-to-noise ratio (SNR) of training HRRPsand test HRRPs are differentpresented Athreshold noise redution method and a recognition model are proposed in the paper to improve the recognition performance in condition of unknown noise Our empirical simulation results show the effectiveness of proposedapproach

Keywords: Radar target recognition, High resolution range profile, Noise reduction.

1 INTRODUCTION

High-resolution range profile (HRRP) can be obtained in high-range resolution radar A HRRP is a one-dimensional image of an object (air target) and represents the projection of the complex returned echoes from target scattering centers onto the radar line-of-sight (LOS) It contains the information about the target’s structure, such as target size, scattering distribution, etc Therefore, radar target recognition using HRRP has drawn researchers’ attention in the automatic radar target recognition area [1-3]

Noise always exists at the output receiver of radar Some of the noise is generated in the transmitter, some of it is cosmic noise and some is added from manufactures sources However, for most radar systems, the majority of the noise is generated in the front end of the radar receiver, particularly by the first amplifier and mixer stages The source of the noise generated in the front end of the radar receiver is the thermal heating of its electrons components, which is better known as thermal noise

The common assumption used in most of the existing statistical models for HRRP data

is that the training HRRPs and test HRRPs are obtained under the similar measurement circumstances In other words, they share the same distribution model and the same statistical parameters In this case, signal-to-noise ratio (SNR) of the training HRRPs and test HRRPs are the same This condition, called known noise, guarantees the optimized recognition performance In real applications, the SNR of the HRRPs is a function of target distance and depends on various factors, such as classes of targets, aspect angles and measurementconditions; thusthe SNR level of test HRRPs usually varies in the progress of observation For known air targets, their training datais usually collected via some cooperative measurement experiments or directly via simulations, which generally has high SNR However, for new (known) air target craft, the prior knowledge (training data)

of which mayonly be obtained using radar measurements, hence the SNR levels of its HRRPs can be low.For this reason, SNR levels of training HRRPsand test HRRPs in the radar target recognition system are usually different This condition, called unknown noise, seriously degrades recognition performance.To solvethis problem, [4], [5] propose statistical models for radar recognition by HRRP Ref.[3] recommends to train neural networks with data of various SNR levels.Drawbacks of these methods are the complexity

of the models and high computational burden.Ref.[1] uses range profiles average methodto raise SNR in HRRP.The effectiveness of this method, however,much depends on the quality of range alignment Another approach is to use a threshold to eliminate noise in HRRP, which is applied in [1], [6]

This paper presents a research of radar target recognition using HRRP in condition of unknown noise and examinesthe effectiveness of the threshold noise reduction methodbased on the threshold equal toh ∗ μ , where μ is estimated mean of the noise amplitude in HRRP, h is coefficient.Section 2 introduces signal-to-noise ratio in HRRP Section 3 presents the proposed threshold noise reduction method Empiricalsimulationresults will be demonstrated in section 4

2 SIGNAL TO NOISE RATIO IN HRRP

HRRP is defined as the vector N complex amplitude values ̇ = [ ̇ , = 1,2 … ]received from N high resolution range cells that radar can receive and separate [7]

In case, in HRRP there is the echo of target class k and noise, we have:

̇ = ̇ = ̇ + ̇ = [ ̇ , = ̇ + ̇ , = 1,2 … ; = 1,2 … ]

where ̇ , ̇ denote the vector N complex amplitude values of target echo and the noise,c is number of target classes of recognition system The echo of one range cell is the sum of sub echoes of scatterers in that range cell

Because elements in HRRP can be considered as uncorrelated, HRRP using in radar target recognition can be expressed by their amplitudes [7] Then, we can express HRRP amplitude under simpler form:

η = , = ̇ ,, = 1,2 … ; = 1,2 … The SNR in HRRP can be defined as:

( ) = 10 × log = 10 × log ∑

∑ η where - average power of the target echo in HRRP;

-average power of noise in HRRP

Amplitude values ; = 1,2 … represent the radar echo intensities from the

corresponding range cells on the target class k( = 1,2 … ) They depend not only on transmitted signal power and target distance, but also on class of target and its aspect angles Thus, the value of , and SNR in HRRP accordingly has high varied properties and uncertainty

3 THRESHOLD NOISE REDUCTION METHOD

There is an evaluation that with the same condition of noise, the SNR of the elements

in HRRP are different The greater the useful signal value in each element is, the smaller the effect of noise on it is If we substruct the amplitude of all elements in HRRP

by using a certain threshold, the SNR in HRRP may increase However, the consequence

of this substruction is that the information of echo elements with amplitude smaller than threshold is lost Therefore, determining the appropriate threshold level is very important

In addition, the threshold must be determined simply and automatically during training and classification stages in order to be able to handle a recognition decision making rapidly

Commonly, in order to reduce the noise we can use the mean value of noise amplitude

in HRRP [1], [6] as a threshold value The determination of the threshold like this is simple and easy to implement With this threshold, however, noise component in HRRP is not significantly removed and recognition quality is not significantly improved because

there are still many noises in HRRP after denoising Generally, the noise amplitude at the receiver output has the Rayleigh distribution Table 1 presents the results of calculations for Rayleigh distribution under various thresholds d, where h is the coefficient with values from 1 to 3, mn is the mean of the random variable

Table 1.The distribution of Rayleigh noise under the thresholdd

It can be seen that when the threshold is moving in the range from 1*mn to 3*mn, the higher the threshold is, the lower the noise is When the threshold is d=3*mn, the noise can

be reduced by 99,9% Thus, to improve the efficiency of threshold noise reduction we need to perform experiments to find out the optimum value of the coefficient h In addition, due to fact that the mean of noise amplitude in HRRP is unknown, there is also a need to build a algorithm to estimate this value

The threshold noise reduction method proposed in this paper uses the threshold ℎ ∗ ̂

to increase the efficiency of noise reduction in HRRP, where ̂ is estimated mean of the

noise amplitude in HRRP, hisa coefficient determined fromexperimentaldata

The algorithm to determine threshold and reducenoise in HRRPis as follows:

- Calculatemean amplitude level over theHRRP: = ∑ ,

- Use m to construct the target mask (threshold of target mask) The amplitude elements of HRRP smaller than m is considered as noise

- Calculate mean noise level ̂ (mean of amplitude elements smaller thanm ): ̂ =

∑ h , , withh , , = 1,2 … is a vector of amplitude elements from HRRP having amplitude smaller than m and is the number of amplitude elements smaller than m

- Used= ℎ ∗ ̂ to reduce noise in HRRP, where h is a coefficient with real value from

1 to 3 (see table 1) The optimum value ofhcan be found out by empirical method by adjusting coefficient h to obtain the highest correct recognition rate of the recognition

system from all the available samples with SNR levels of interest

4 SIMULATION RESULTS AND DISCUSSION

The main objectives of these experiments are to evaluatethe quality of radar target recognitionusing HRRP in conditions of unknown noise and to examine the effectiveness

of the proposed noise reduction algorithm bysimulation in Matlab with the radar recognition flowchartshown in figure 1

The radar target backscattering simulation program (RTBS) [8] were used to create complex HRRPs of 9 target classes with the radar mode set as [3] The HRRP is a 120-dimensional vector The target class names are Tu16, B1b, B52, Mig21, Tornado, F15, Alcm, Glcm and Decoy For each of 9 target classes, 900 HRRPs are thus generated corresponding to 900 azimuthal angular positions, from 0 to 180, with angular increment

of 0,2 [9] and elevation angle of 3° The 900 HRRPs of each target class were divided into a training data set and a test data set Atraining data set was created by uniformly

sampling the aspect angle per every0,4 Therefore, 450× 9 = 4050 HRRPs were used to

createatraining database and the remaining 4050HRRPs were used for classification (test data)

Figure 1 Radar recognition flowchart

To simulate the effect of noise, complex additive white Gaussian noise was added to bothtrainingand test data to achieve the desired SNR The amplitudes ofcomplex noisy data was calculated to form the real data Data were preprocessed using 2 steps: denoise (using the proposed noise reduction method) and normalization of each HRRP, i.e.,

||HRRP|| = 1

The RBF network classifier:radial basis function is Gaussian function j( ) =

,where x is input data, m and are the center and the width parameter; the number

of hidden layer neuron is 90; thetraining procedure is thetwo-phase learningper [10] The RBF network classifierwas trained by training data set with SNRtaking the value

of 0dB, 5dB, 10dB, 15dB, 20dB or 25dB in succession The recognitionwas performed for each SNR of training data setby test data set with SNR changedsuccessively between the

values mentioned above The average correct recognition rate(ACRR) Pc was used as the

classification result

Achievedresults were depicted in figures 2 with 6 experiments corresponding to 6 different SNR levels of training data set (Experiment1Experiment 6).These experiments arerepresentativesused tostudy the condition of unknown noise The experiments 16 correspond to SNR levels of training data setof 25dB, 20dB, 15dB, 10dB, 5dB, 0dB respectively.In each experiment, we give the average correct recognition rate (ACRR) for two cases “Denoise” and“Non-denoise”withvariousSNR levelsof test data “Denoise” denote using the proposed noise reduction method with h=1.0 and h=1.8 and “Non-denoise” denotes not using noise reduction.“Non-“Non-denoise” case shows clearly that in the absence ofnoise reduction, recognition performance dependssignificantly on the difference between SNR levels oftraining and test data The bigger the difference is, the worse the recognition performance is, even with high SNR level of test data.For example, in experiment 6, we have the highest average correct recognition ratewhen the SNR of test datais 0dB; when the SNR of test data increases (i.e the difference between SNR levels of training and test data becomes bigger),the recognitionperformancedecreases.“Denoise” cases show that when using a proposed noise reduction algorithm, the recognition performanceswere improved significantly in all experiments The result supports the statement that the proposed noise reduction algorithm has improved the SNR in HRRP and reduced theeffect of noise on the correlationbetweenHRRPs, thus brings higher recognition performance, especially when there is considerable variance between SNRs of the training HRRPs and the test HRRPs and when SNR of test data is low

Training data

RBF Network classifier Test data

Noise

Pre-Processing +

Pre-Processing

Figure 2.The recognition performance for different noisy cases of training data

When we changed the value of h, the simulation results showed that the recognition performance is improved the most when h is chosen between 1,8÷2,4at low SNR of test dataand this is significantly better than in the case of selecting h = 1 as in [1], [6] (see figure 2)

From these experiments, we can remark that the classifiers can well recognize the HRRP with SNR levelclose to SNR level of training data Hence, a useful and practical approachto improve the recognition performance in condition of unknown noise is thatwe use the recognition model with two parallel clasifiers [called 2-SNR parallel classifiers

model: two-SNR PCM], where the first classifier is trained by training data set with high

SNR (for example 20dB) and the second by low SNR (for example 5dB) The classifier trained by low SNR level is used to recognize targets at long distance or HRRPs with low SNR and the classifier trained by high SNR levels is for targets at short distance The average correct recognition rates for this model (using the proposed noise reduction

method) for different SNR levels of test data were shown in figure 3 (curve two-SNR PCM), where the first classifier recognized test data with SNR from 0dB to 14dB, and the

second- from 14dB to 30dB The superiority of our proposed approach can be seen clearly

when we compare our results with the results obtained by [4]: the curve two-SNR PCMis

very close to the “known noise” curve

0

0.5

1

Experiment1 (25dB)

SNR of test data (dB)

0 0.5

1 Experiment2 (20dB)

SNR of test data (dB)

0 0.5

1 Experiment3 (15dB)

SNR of test data (dB)

0

0.5

1

Experiment4 (10dB)

SNR of test data (dB)

0 0.5

1 Experiment5 (5dB)

SNR of test data (dB)

0 0.5

1 Experiment6 (0dB)

SNR of test data (dB)

Non-Denoise Denoise(h=1) Denoise(h=1.8)

Figure 3.The recognition performance for two-SNR PCM and “known noise” case

5 CONCLUSION

In this study, we considered the radar target recognition using HRRP in condition of unknown noise.The experimental results show that:

- It is necessary to reduce noise in HRRP at the preprocessing stage in radar target recognition system Using the proposed threshold noise reduction algorithm can raise SNR

in HRRP and significantly reducethe effect of noise on recognition performance, especially when there is considerable variance between SNRs of the training HRRPs and the test HRRPs and when SNR of test data is low This algorithm is simple, implemented automatically for each HRRP and hasshort processing time The optimal threshold can be found out by empirical method by adjusting coefficient h for each recognition system

- The proposed two-SNR PCM model in this paper can handle the problem of radar target recognition in condition of unknown noise The simulated results show that thismodel has the recognition results which are approximately equalto which of known noise case This project is quite simple and can be implemented in real radar target recognition system using HRRP

REFERENCES

[1] A Zyweck, R E Bogner,"Radar target recognition using range profiles",IEEE

International Conference, Adelaide, 19-22 Apr 1994,Vol 2, pp 373-376, (1994)

[2] S Kent N G Kasapoglu and M Kartal, “Radar target classification based on

support vector machines and high resolution range profiles”,IEEE Radar Conference,

Rome, 26-30 May 2008, pp 1-6, (2008)

[3] Д Ф Ле, “Распознавание радиолокационных целей по дальностному портрету

с использованием aппарата нейронных сетей в целях мониторинга воздушного пространства” Диссертация на соискание ученой степени

кандидата технических наук, Санк-Петербург, (2006)

[4] H.Liu, L.Du, P.Wang, M.Pan, Z.Bao, “RadarHRRPAutomaticTargetRecognition:

AlgorithmsandApplications”, IEEE InternationalConferenceon Radar, Vol 1,

Chengdu, 24-27 Oct 2011, pp 14-17, (2011)

[5] Х H Нгуен, “Структурно - параметрическая адаптациярешающегоправила системы

распознавания – измерения",Диссертация на соискание ученой степени кандидата

технических наук,Минск, (2003)

[6] L Xiankang, G Meiguo, F Xiongjun, “Satellite Recognition Base on Wavelet

Denoising in HRRP Feature Extraction”, 2007 2nd IEEE Conference on Industrial Electronics and Applications, Harbin, 23-25 May 2007,pp 2530-2533,(2007)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

SNR (dB)

Known noise Two-SNR PCM

[7] Охрименко А Е, “Основы извлечения обработки и передачи информации Часть3:

Распознавание – различение”, Мн БГУИР, (1994)

[8] S A Gorshkov, S P Leschenko, V M Orlenko and Y D Shirman, “Radar Target

Backscattering Simulation Software and User's Manual”, Boston, Artech House,

(2002)

[9] C R Smith and P M Goggans, “Radar target identification.”, IEEE Antennas and

Propagation Magazine, April 1993, Vol 35, pp 27-38, (1993)

[10] C M Bishop, “Neural networks for pattern recognition”, Oxford, Clarendon Press,

(1995)

TÓM TẮT

NHẬN DẠNG MỤC TIÊU RA ĐA THEO CHÂN DUNG CỰ LY

TRONG ĐIỀU KIỆN KHÔNG BIẾT NHIỄU

Bài báo trình bày kết quả nghiên cứu mô phỏng bài toán nhận dạng mục tiêu ra đa theo chân dung cự ly phân giải cao trong điều kiện không biết nhiễu – là điều kiện khi

mà tỷ số tín trên tạp của các chân dung cự ly huấn luyện và nhận dạng khác nhau Một phương pháp giảm nhiễu theo ngưỡng và một mô hình nhận dạng được đề xuất trong bài báo nhằm nâng cao chất lượng nhận dạng trong điều kiện không biết nhiễu Kết quả mô phỏng đã cho thấy hiệu quả của hướng nghiên cứu được đề xuất

Từ khóa: Nhận dạng mục tiêu ra đa, Chân dung cự ly, Giảm nhiễu

Author affiliations:

1

Academy of Military Science and Technology;

2

Military Technical Academy;

*Corresponding author:ngthanhhungvn@gmail.com