CC BY
10
ТЕХНИЧЕСКИЕ НАУКИ
AN ANALYSIS OF THE STANDARD DEVIATION OF THE ENTROPY PARAMETER FOR TARGET _DETECTION ON THE BACKGROUND CLUTTER_
DOI: 10.31618/ESU.2413-9335.2021.6.84.1305
Pham Trong Hung
Ph-D, Military Technical Academia, Vietnam Republic
Nguyen Tien Tai
Ph-D, Military Technical Academia Vietnam Republic
Nguyen Van Hai
Ph-D, Military Technical Academia Vietnam Republic
ABSTRACT
The paper proposes a method to improve the performance of target detection on the background clutter using the standard deviation of polarization entropy H parameter. The performance of target detection is analyzed with different types of target, on the background clutter models Rayleigh, Weibull and Laplace. Also examined the special case when the target has the same H parameter as the background clutter. The results showed that the performance of the method using the standard deviation of H in the target detection on the background clutter has increased significantly compared to the case where only the H parameter is used.
Key words: standard deviation, target detection on the background clutter, polarization entropy.
1. Introduction
Because of the strong fluctuation and the inhomogeneous property of the background clutter, the detection of small sea targets has always been a problem with sea radar systems, even with the detection based on the Doppler effect. In this situation, polarimetric information could be a useful and important tool to improve the detectability.
Since the optimum detector was proposed by Novak 1989 [1], several polarimetric algorithms have been developed [2], [3], [4]. Most of them apply general likelihood ratio test (GLRT) [5], under the assumption that covariance matrix of the background clutter can be calculated by training data, and this covariance matrix is considered as prior knowledge. The detection performances of those detectors, however, decrease significantly in inhomogeneous and non-stationary clutter environment [6].
In order to address this problem, paper [1] proposed the use of the complex Gaussian distribution to model the inhomogeneous property of the background clutter. However, the use of this non-Gaussian distribution model increases the complexity of the optimum detection algorithm. For example, the detection statistic based on the proposal scheme presented in [7] has a complete theoretical expression only for a special case of two polarimetric channels, and the detector proposed in [8] does not support the constant false alarm rate (CFAR) capability for the background clutter.
Another approach to solve this problem is to use the degree of polarization, DoP for the detection, as suggested in [9]. This method is under investigation due to a few unreasonable results. In [10], the authors
proposed the use of the Weighted Average H (WAH) and the Weighted Average a (WAa). The test results with real data showed that this method has a high false alarm rate.
In this paper, the author will analyze and investigate the standard deviation of the entropy parameter in order to detect the target on the background clutter.
The remainder of this paper is organized as follows: section 2 reviews the methodology needed to obtain the entropy H of the polarization covariance matrix. Sections 3 proposes a test procedure to detect target on the background clutter based on the entropy H and the standard deviation of H. Simulation results of the target detection on the background clutter of the proposed method and compared the detection efficiency using the entropy parameter and the standard deviation of H presented in Section 4, and the conclusion is provided in Section 5.
2. The scattering entropy
The scattering matrix S = Shh Shv is normally
iSvh ^vv
used to describe the relation between transmitting and backscattering fields. With common monostatic radar we have shh svh. The received signals are range-processed and integrated to form an estimate for the 3-dimensional Pauli scattering vectors [11] at each range bin:
ki — 1[(shh + svv) (shh-svv) 2 shv F (1)
The 3x3 dimensional polarimetric coherency matrix can be calculated as:
N
(\T3\)=1^kik{
i\shh + svv\2) ((shh + svv) (shh — svvY) (2(shh + svv) shv) T11 T12 T13-
((shh — svv) (shh + ^vv) ) d^hh ^vv \ ) (2(shh + svv) shv) = T21 T22 T23
i2shv(shh + svvY) (2shv(shh — svvY) (4\ShV\2) IT31 T32 T33-
The eigenvalues of the T3 matrix can be calculated from the equation:
T?-AI = 0
(3)
where I is the unitary matrix and 1 is the eigenvalue of the T3 matrix.
Here introduced the definition of polarimetric entropy (H), proposed by Shane Cloude [11]. The polarimetric entropy (H) of wave is obtained from the knowledge of the two eigenvalues I1, I2 and I3 of the coherency matrix T, and difined as:
H = -Y.UPk log3(pk)
withpfc =
\k
X1+X2+
(4)
(5)
H assumes any real value between 0 and 1. If H = 0, the measured environment is so-called perfectly polarized. When H=1, the measured environment is so-called completely unpolarized and presenting "polarimetric white noise".
3. The detection test
In this section, the authors develop the detection test. The question here is to determine if a target is present within CUT (Cell under Test) or not, based on the data received. In other words, the detection problem consists of choosing one from two possible hypotheses: hypothesis Ho (no target presence) and hypothesis Hi (target presence)
fro(H),Ho rl(H),H1
(2)
(6)
The optimal detector applies the likelihood ratio test [1], in that the optimal detector produces the maximum probability of detection (Pd) with the given false alarm rate (Pfa). In this case, however, there is no prior knowledge about the data distribution. As a result, we can not apply the likelihood ratio test in this problem. Another method is the GLRT, in which the unknown parameters of the data distribution are replaced by their maximum likelihood estimated values [12].
Similarly, when using the standard deviation of the H parameter for the detection problem, the problem arises to decide whether or not the target in the CUT based on the received data. On each radar cell corresponding to each received signal from the polarization channels, the corresponding H will be calculated according to Equation (4). To calculate the standard deviation of H it is necessary to have N values of H. If oH < 0™ it decides to have a target in that radar cell and vice versa. Flow chart of detection algorithm using H and oh as shown in Figure 2.
Ho'.
Hi-
-Th
°H < VHh
(7)
where a™ is the detection threshold by oh and can be calculated by desired Pfa in the case where only background clutter is present.
1=1
T
H
Begin
Ehh, Ehy, Evh, Ew; PF
1
Count detection
threshold rT ' H
1
Count TH
Target presence
End
Target presence
End
a b
Figure 2. Algorithm for target detection using H and aH parameter
4. Analysis of the detection performance 4.1. Target detection on the background clutter using the entropy H and standard deviation of H with different targets
In this section, using the background clutter with Rayleigh distribution, creating 3 target types with different entropy parameters are: 0.76; 0.88; 0.92. The results are presented as shown in Figure 3. Figure 3a describes the entropy parameter obtained with two cases: only background clutter and background background clutter plus target 1.
Figure 3 a shows the entropy values for background clutter fluctuating over a wide area « 0 + 1, meanwhile with target 1 (H = 0,76) plus background clutter (figure below), the entropy parameter fluctuates in the range of « 0,6 + 1. In this case, the entropy value of the target plus background clutter is also not equal to the actual entropy value of the target (H = 0,76). Figure 3b is the dependence of H for the target plus background clutter with the signal to clutter ratio (SCR). On this figure shows the mean value of entropy, H « 0,57 for the background clutter. Meanwhile, the mean entropy value of the target plus background clutter increases as SCR increases. The higher the SCR, the closer the target's average entropy plus the background clutter will get closer to the actual
entropy of the target. Target 1 with minimum entropy (H « 0,76) is on the bottom, target 3 with the higher antropy (H « 0,92) is on the top. Thus it can be seen that there is a difference of the entropy parameter in the case where the target is on the background clutter compared to the case with only background clutter. This is the signature which can be used the entropy parameter for the target detection on the background clutter.
Figure 3 c shown the standard deviation of entropy for the above targets. Figure 3 c shows that the standard deviation of H for background clutter has a higher value and is approximately 0.25. Whereas the standard deviation of H for the background clutter plus target is smaller than for the case with only background clutter. As the SCR increases, the standard deviation for the target case plus background clutter decreases, reaching 0. Thus, we see a large difference in the standard deviation of the background clutter plus target case compared to one there is only background clutter. While the standard deviation of H for the background clutter is stable, the standard deviation of H for the background clutter plus target changes and decreases as the SCR increases, so this standard deviation can also be used in the detection problem (differentiating) the target on the ground surface.
SCR(dB|
c)
d)
Figure 3.
Figure 3d shows the results and the comparison of the target detection performance with parameter H and standard deviation of H on the 3 above target types using Monte Carlo simulation method. The number of samples is 12,000. Probability of false alarm in this case is Pf = 10-5. The threshold of detection is set so that in the case of target plus background clutter, the ratio of the number of samples with H less than the threshold Y™ and &Hh to the total number of samples equals to Pf. From Figure 3d, it can be seen that, in the case of using the entropy parameter, the PD for target 3 is the highest and for target 1 the lowest. Specifically with SCR = 0 dB, the Pd for target 3 equals 0.85; for target 2 it equals to 0.78; and for target 3 it is equals to 0.6. This means that, as further away the entropy of the target from the entropy the background clutter, the higher the probability of target detection Pd. When comparing the method using H and oh, it is found that the probability of detection Pd using gh is higher than using H parameter. For example with SCR = 0,PD « 1 and SCR does not have much influence on Pd.
4.2. Target detection using H and gh in the case where the target and the background clutter has the same entropy value
Another advantage of using the standard deviation of the entropy H in the target detection on the background clutter is that when the target has the entropy H close to the average entropy of the
background clutter, the results are presented as shown in figure 4a. In this case it is not possible to distinguish the target on the background clutter using H parameter there is no difference of this parameter in two cases: background clutter only and background clutter plus target. The modulation was tested with different types of background clutter models and a target withH « 0,59. The mean value of entropy H and the standard deviation of H for the different models of background clutter are shown in Figures 4b and 4c.
Figure 4b shows that the mean value of entropy H of the background clutter and the background clutter plus the target with different clutter models are same and equals 0.5. Thus, based on this parameter it will be impossible to distinguish the target on the background clutter. However, considering the results shown in Figure 4c it is found that the standard deviation of the entropy parameter in the case of the background clutter plus target is different from the case where only the background clutter. The value of the standard deviation in that case decreases as the SCR increases. Thus, it can be seen that, when it is not possible to use the entropy parameter for the target detection, it is possible to use the standard deviation of the entropy parameter to detect the target on the background clutter. The performance of the detection for this type of target is be analyzed. The results presented in Figure 4c with different models of background clutter.
a)
—h- Clutter Raylelgh -е—Tar+CI.Weibull Tar+Cl-Uplsce -----Тэг+CI Lgplgcs
l"J. ■ II-1 I—l-H H—*t"4—+-H I—hl I 4—I----h 1Ч--Н
N„=0.59 Н,г=0.5Э H -0.59
-10 -5 0 5 10 15 SO
SCR (dB)
Cj
Hinh 4
Figure 4d shows that it is impossible to detect a target using the entropy parameter while if using the standard deviation of this parameter, the probability of target detection is high, for example with SCR = 5 dB, the Pd=1 for different models of background clutter when using oh. This is the most obvious effect of using the standard deviation entropy parameter in target detection problem. Regardless of whether the entropy parameter of the target is the same as the entropy parameter of the background clutter, the target can be detected on that background clutter using the standard deviation of it.
5. Conclussion
The paper proposed a method of target detection on the background clutter using the standard deviation of the polarization entropy H. The results of the performance detection using entropy H and oh with different target types, different models of background clutter and the special cases when the target has the same H parameter as the background clutter. The results show that the performance of the method using the standard deviation of entropy parameter is better in all cases than using the entropy parameter only. Especially, in the case where the entropy parameter of the target is identical to that of the background clutter's entropy, it is not possible to use the entropy parameter for the detection problem, but the target can be detected using the standard deviation of that parameter. This new method can be applied to other polarization parameters when detecting or distinguishing targets on the background clutter according to a certain polarization parameter.
References
SCR (dB) b)
Q. 0.5
SCR [dB]
[1]
[2]
[3]
[4]
[5]
[6]
[7]
d)
Novak L. M. en Sechtin M. B., „Studies of target detection algorithms that use polarimetric radar data," IEEE Trans. on Aerosp. Electron. Syst, vol. 25, nr. 2, Mar. 1989., pp. 150-165, 1989. Park H. R, Li J en Wang H, „Polarization-space-time domain generalized likelihood ratio detection of radar targets," Signal Processing, vol. 41, p. 153—164, 1995.
Pastina D, Lombardo P en Bucciarelli T, „Adaptive polarimetric target detection with coherent radar. Part I: Detection against Gaussian
background," IEEE Trans. on Aerosp. Electron. Syst, Vols. %1 van %237, No. 4, pp. 1194-1206, 2001.
Hurtado. M en Nehoira. A, „Polarimetric detection of target in heavy inhomogenous clutter," IEEE Transactions on Signal Processing, vol. 56, nr. 4, pp. 1349-1361, 2008.
Kelly E. J, „An adaptive detection algorithm,"
IEEE Transactions on Aerospace and Electronic Systems, AES-22, vol. 1, p. 115—127, 1986.
Park. H en Wang. H, „Adaptive polarization-space-time domain radar target detection in inhomogeneous clutter enviroments," Inst. Elect. Eng. Proc. Radar Sonar Navig., vol. 153, pp. 3543, 2006.
Lombardo. P, Pastina. D en Bucciarelli. T, „Adaptive polarimetric target detection with coherent radar. Part II: Detection against non-
Gaussian background," IEEE Trans. Aerosp. Electr. Syst, vol. 37, pp. 1207-1220, 2001.
[8] De Maio.A en Alfano. G, „Polarimetric adaptive detection in non-Gaussian noise," Signal Processing, vol. 83, pp. 297-306, 2003.
[9] Bo Ren, Longfei Shi en Guoyu Wang, „Polarimetric Target Detection Using Statistic of the Degree of Polarization," Progress In Electromagnetics Reserch M, vol. 46, pp. 143152, 2016.
[10] Peng Wu, Jun Wang en Wenguang Wang, „A Novel Method of Small Target Detection in Sea
Clutter," International Scholarly Research Network ISRN Signal Processing, vol. 33, nr. 4, pp. 816-822, 2011.
[11] Cloude S.R en Pottier E, „An entropy based classification scheme for land applications of polarimetric SAR," IEEE Transactions on Geoscience and Remote Sensing, vol. 35, nr. 1, pp. 68-78, 2008.
[12] S. Kay, Fundamentals of Statistical Signal Processing: Detection Theory, NJ: Prentice-Hall: Englewood Cliffs, 1993.
РАЗРАБОТКА ИНТЕЛЛЕКТУАЛЬНОЙ СИСТЕМЫ РАСПОЗНАВАНИЯ И ИДЕНТИФИКАЦИИ
В РЕЖИМЕ РЕАЛЬНОГО ВРЕМЕНИ
]Боранбаев С.Н., 2Амиртаев М.С.
1,2 Евразийский национальный университет имени Л.Н. Гумилева,
Нур-Султан, Казахстан
АННОТАЦИЯ
В этом исследовании была построена система для распознавания лиц в режиме реального времени с применением инструментов Open Face библиотеки Open CV. В статье приведена методология создания системы и результаты ее тестирования. Библиотека Open CV имеет различные модули выполняющие множество задач. В данной работе модули Open CV были использованы для распознавание лиц на изображениях и идентификации лиц в режиме реального времени. Кроме того, метод HOG был применен для того, чтобы обнаружить человека по передней части его лица. После выполнения метода HOG, 128 измерений лица были получены с помощью метода кодирования изображений. Затем была использована сверточная нейронная сеть для идентификации лиц людей с помощью алгоритма линейного классификатора SVM.
1. Введение
Быстрое развитие технологий влияет на потребность в системах распознавания. Одной из быстро развивающихся технологий является распознавание лиц. Технологии распознавания лиц являются частью компьютерного зрения. Множество исследователей были посвящены разработке системы распознавания лиц из-за ее широких возможностей применения в коммерческой и государственной сферах. Технология распознавания лиц может применяться в различных областях, таких как системы наблюдения, системы аутентификации, поиск людей и др.
Есть ряд причин, по которым процесс распознавания лиц является очень затруднительным. К примеру, объект в виде лица человека может быть различным из-за разных форм лица и цвета кожи, а также другие объекты могут частично накладываться на лицо. Возможности камеры идентифицировать лицо часто зависит от качества видеозаписи и уровня освещения. На данный момент, некоторые обстоятельства (такие как качество камеры, освещение, угол обзора лица и т.д.) не позволяют достичь соответствующей степени распознавания в изображениях и видеопотоках, они же влияют и на качество идентификации.
2. Система классификации и распознавания человеческого лица
2.1 Поиск всех лиц на изображении. Сделать поиск лица на изображении необходимо для того, чтобы выбрать область изображения, которая передается на дальнейшую обработку. Для этого используется метод HOG (гистограмма направленных градиентов) [1].
Согласно алгоритму, изображение преобразуется в черно-белый формат, поскольку для поиска лиц не нужны цветовые данные. Алгоритм смотрит на каждый пиксель и соседние пиксели, чтобы узнать, насколько темным является текущий пиксель по сравнению с окружающими пикселями. Затем добавляется стрелка, указывающая, в каком направлении изображение становится темнее. После выполнения этой процедуры для каждого отдельного пикселя изображения каждый пиксель заменяется стрелкой. Эти стрелки называются градиентами, и они показывают направление от светлых к темным пикселям по всему изображению. Если вы используете темные и светлые изображения одного и того же человека, пиксели будут иметь разные значения яркости, но если учесть направление изменения яркости, вы получите одно и то же изображение независимо от яркости исходного изображения. Чтобы сэкономить ресурсы и избавиться от избыточной информации, изображение делится на квадратные блоки размером 16х16 пикселей. Затем делается замена каждого такого квадрата на изображении стрелкой,
