Você está na página 1de 11

Multilizer PDF Translator Free version - translation is limited to ~ 3 pages per translation.

Artigo
Identication de objectos enterrados em GPR
usando Amplitude modulada sinais
extrados de anlise multiresoluo
monognicas sinal
Lihong Qiao 1,2 , Yao Qin 1,2, *, Xiaozhen Ren 1,2 e Qifu Wang 3

Recebido: 31 de julho de 2015; Aceitos: 25 de novembro de 2015; Publicado em: 4 de dezembro de 2015
Editor acadmico: Assefa M. Melesse
1 Faculdade de cincia da informao e engenharia, Universidade de Henan da tecnologia, 450001 de
Zhengzhou, China; lhqiao@Haut.edu.cn (QI); rxz235@163.com (X.Z.)
2 Chave laboratrio de processamento de informaes de gro e controle, Henan
Universidade da tecnologia, Ministrio da educao, Zhengzhou 450001, China
3 Henan Academia de cincia, aplicada fsica Instituto co., Ltd, Zhengzhou 450001,
China; ewangqifu@hotmail.com
* Correspondncia: eqinyao@163.com; Tel.: + 86-27-6775-6529

Resumo: necessrio detectar o alvo refiections em imagens de radar (GPR) penetrante de


cho, para que a superfcie alvos de metal podem ser identificam com xito. Para localizar
com preciso enterrado
objetos de metal, um novo mtodo chamado o sistema de anlise de sinal monognicas
multiresoluo (MMSA) aplicado em imagens de radar (GPR) penetrante de cho. Este
processo inclui quatro etapas. Primeiro a imagem decomposta pelo MMSA para extrair o
componente da amplitude da imagem B-scan. O componente de amplitude aumenta a refiexo de
alvo e suprime a onda direta e refiective onda a uma extenso grande. Depois usamos a regio do
mtodo de extrao de interesse para localizar o alvo genuno refiections de esprias
refiections calculando a varincia normalizada do componente amplitude. Para encontrar que os
pices dos alvos, um Hough transformam usado na rea restrita. Finalmente, calculamos a
posio horizontal e vertical do alvo. Em termos de deteo de objeto enterrado, o sistema
proposto apresenta desempenho promissor, conforme os resultados experimentais.

hiprbole de refiexo; GPR; anlise multiresoluo monognicas sinal; amplitude


Palavras-chave:
componente; Transformada de Hough

1 Introduo
Como um mtodo geofsico, radar de Prospeco Geotcnica (GPR) usado na
deteco dos destinos prximos s reas de superfcie de solo para localizar objetos
subterrneos tais como minas terrestres e tubos. Focalizamos a deteo de metal ponto neste
trabalho. Hiprboles todos tem que ser localizado e cada nico radargram tem de ser analisado,
com a finalidade de extrair a posio exata dos alvos.
eoutros.
Existem algumas pesquisas na deteo de objeto enterrado. Para identificar a hiprboles, Delbo [1]
propor um procedimento baseado em wavelets para reduzir o rudo e, em seguida, uma
abordagem fuzzy cluster para detectar objetos. Articial redes neurais so usadas em [2]. Um
processo de reconhecimento de padro usado para localizar objetos no subsolo atravs de um
processo de reconhecimento de padro [3,4]. Interpretao em tempo real de imagens GPR
proposta em [5].
Um modelo chamado multiresoluo monognicas sinal anlise (MMSA) foi
desenvolvido Vrpor Unser [6].
ios componentes tais como componentes de amplitude e fase so obtidos
atravs da anlise do sinal monognicas, que tem uma apresentao adequada de
multiresoluo. Os componentes baseados em wavelet amplitude suprimem o rudo e melhorar a alvos em imagens GP
eoutros.
[7] propor
abertura sinttica radar (SAR) alvo classication mtodo baseado na geometria de Riemann. Combinado

Sensores2015, 15, 30340 30350; doi:10.3390 / s151229801 www.mdpi.com/Journal/Sensors

Multilizer PDF Translator Free version - translation is limited to ~ 3 pages per translation.
Multilizer PDF Translator Free version - translation is limited to ~ 3 pages per translation.

Sensores2015, 15, 30340 30350

com uma matriz de covarincia, os componentes monognicas so usados em classication de


imagem SAR. Alm disso, prope-se a representao esparsa do sinal monognicas de
imagens SAR em [8].
Este trabalho tem como objetivo apresentar um novo mtodo para resolver as questes de
deteo de objeto. O mtodo inclui pr-processamento dos dados usando um MMSA para
extrair o componente de amplitude. Aps a extrao, desordem de fundo removido em
grande medida, mantendo o sinal do alvo, que visa livrar-se dos alvos falsos e localize os alvos
com preciso. Para limitar o escopo de refiexo hiprboles para determinadas reas, usamos
o mtodo de extrao de ROI em dados GPR, que diminui a complexidade computacional de
hiprbole tting. Em seguida, a transformada de Hough usada para tting de hiprbole.
Finalmente, calculamos a posio horizontal e vertical do alvo. Experimentos em imagens GPR
provm do mtodo do domnio do tempo (FDTD) nite-diferena e situaes realistas.

O livro est organizado da seguinte forma: o modelo de refiexo de um objeto enterrado


apresentado na seo 2. O algoritmo de deteco baseado em MMSA, extrao de ROI e
Hough aplicativo transform apresentado na seo 3. Os resultados experimentais so
ilustrados na seo 4. Finalmente, as concluses so apresentadas na seo 5.

2 Modelo de refiexo de um objeto enterrado


Um sistema de radar contm um par de antena transmissora e receptora. Primeiro, uma reviso
do modelo de refiexo de alvo ponto metal enterrado. Quando a antena de terra-acoplado examina
um objeto linearmente, a posio da antena
x e o correspondente atraso de tempo tdo
aproximadamente,
eco satisfazer um trs-parmetro
equao da hiprbole:
t2 4px x0 q2
1 (1)
t20 2 2
VT
Sensores2015, 15 0 3
Onde v a velocidade de propagao de onda subterrnea, considerado constante em uma pequena regixo0,
a posio horizontal do objeto, e
posio de objeto do vrtice. A velocidadev e a0posio
a posio de objeto do vrtice. A velocidade
(x , t ) do vrtice determinar a forma de
t o tempo de atraso para o eco. O pice px0 , t0 q indica
v e a posio 0 0 px0 , t0 q do vrtice determinar o
a hiprbole
formadedareflexo.
hiprbole Um modelo
refiex simplessimples
o. Um modelo de reflexo para
de refiex GPR
o de GPR mostrado
mostradonanaFigura
Figura1.1.

x x0
Overground

Underground ( X0 , )T0

antena
objeto
( X, )T

Figura 1. Um modelo simples de refiexo de GPR. A posio da antenax e o atraso de tempo do eco
Figura 1. Um modelosatisfazer
t aproximadamente simplesade reflexo
equa x e o tempo de eco
para(1).GPR. A posio da antena
o hiperblica
demorat aproximadamente satisfazer a equao hiperblica (1).
3 Algoritmo da deteo

3. detecoOalgoritmo
mtodo de deteo de objeto consiste em quatro etapas:

1 Extrair os componentes de amplitude, usando o mtodo MMSA;


O mtodo de deteo de objeto consiste em quatro etapas:
2 Extrair a regio de interesse para estreitar a regio de determinadas reas;
1 Extrair
3 os componentes
Localize de amplitude,
hiperblicos padr es atravs deusando o mtodo
uma transforma MMSA;
o de Hough;
2 4 Estime a posi o horizontal e vertical do alvo.
Extrair a regio de interesse para estreitar a regio de determinadas reas;
3 Localize hiperblicos padres atravs de uma
4 transformao de horizontal
Estime a posio Hough; Oeowchart
vertical domostrado
alvo. na
Figura 2.
O fluxograma mostrado na Figura 2.
30341
Dados brutos
Multilizer PDF Translator Free version - translation is limited to ~ 3 pages per translation.
1 Extrair os componentes de amplitude, usando o mtodo MMSA;
2 Extrair a regio de interesse para estreitar a regio de determinadas reas;
Multilizer
3 PDF Translator
Localize Free
hiperblicos version
padres - translation
atravs is limited
de uma transformao to ~ 3 pages per translation.
de Hough;
4 Estime a posio horizontal e vertical do alvo.
Sensores2015, 15, 30340 30350
O fluxograma mostrado na Figura 2.

Dados brutos

Estou extrao usando a multiresoluo


Anlise de sinal monognicas

Extrao de regio de interesse

Transformada de Hough

Estimando o horizontal e vertical


posio do alvo

Figura 2. O fiowchart da estratgia proposta.


Figura 2.O fluxograma da estratgia proposta.
No detalhe, a imagem GPR primeira decomposta pela transformao wavelet monognicas para extrair
o componente de amplitude. O componente de amplitude aumenta a refiexo de alvo e suprime
em detalhe, a imagem GPR primeiro decomposta pela transformao wavelet monognicas
para extrair a onda direta e a onda de refiective, em grande medida. A informao de fase
transformada
decomposto wavelet
pela monognicas
oreflexiva
componente de cont m muita
amplitude. informa o de textura detalhada de todas
onda
as direta
ondas e edifcil
ondalocalizar oem grande
alvo. O m medida.
t odo AOinformao
pode
componente
n o s dede
intensificar
amplitude
fase
o decomposta
alvo, mas
aumenta
tamb
a reflexo de
pelamtransformao
alvo e suprime a
wavelet monognicas
eliminar os indesejveis contm muita informao de textura detalhada de todas as ondas e
presena do eco de superfcie do solo. Em seguida, usamos o mtodo de extrao de ROI para
restringir
um pequenoa regi o a difcil
intervalo. A de localizar
extra o alvo.
o baseia-se O mtodo
a varincia pode no s maior
normalizada intensificar
que um o alvo, mas
nmero de
tambO
xed. m custo
eliminar os indesejveis
computacional da transformada de Hough significativamente reduzida. Ento
a transformao de Hough
usado para detectar padres de hiperblicos.

3.1.PreprocessingBasedontheMultiresolutionMonogenicSignalAnalysisStructure
O MMSA mtodo os Riesz-Laplace wavelet transformaes.Trs wavelet
componentes amplitude componentes fase componentes e instantneo frequncia
de usos,
componentes so extrados. Esses parmetros tm uma denition muito clara em rdio frequncia
teoria e eles podem ser estendidos para dois sinais dimensionais. Usando o componente de
amplitude, desordem de fundo removido em grande medida, mantendo o sinal do alvo.
Nesta seco ns primeira reviso o sinal monognicas. Em seguida, apresentamos as
transformaes de wavelet de Riesz-Laplace para implementar o mtodo MMSA e obter as
caractersticas de modulao, que incluem a amplitude, fase e frequncia.
3-componentes do sinal monognicas f pq so dened em [9] como:

f m pq
f , Re pf< , Imp< f ppqqqq
(2)
p f , f1, f2q

Onde f pq um sinal dimensional, p x, yq, e < a transformao de Riesz de


f pq .
Componente de amplitude dened por:
b
Apxq x p q||
||f m f 2 ` f 12 ` f 22 , (3)

A orientao e fase so dadas por:


f Acos, f 1 APorque,
of 2pecado
Apecado
pecado (4)

30342
Multilizer PDF Translator Free version - translation is limited to ~ 3 pages per translation.
Sensors 2015, 15, 3034030350

In particular, the amplitude components and phase components are recovered in the direction
u p cos, sin qp fr{
1 , fr{
2 q , which can be extended to other directions.
Then we show how to realize a monogenic representation of wavelet coefficient. On each
bandpass-filtered signal:
p f i qpxq (5)

detp Dq|i{2 Dxi is the normalized and spine wavelet.


`
x
where i pq|
Due to the symmetry characteristics, the coefficients of the signal are defined as:
@ D
i rks f , i,k = pi f qpq|
x xDpi`1q k (6)
@ D @ D
r1,i rs`
k jr2,i rs
k f , < i,k =< f i,k xDpi`1q k (7)

where Riesz coefficients real component is r1,i rks and the imaginary component is r2,i rks. k is the space
index and i is the scale index.
Based on previous knowledge, the MMSA is applied to extract the amplitude components in
the wavelet-domain. The tensor method is used to estimate the local orientation. The
corresponding directional Hilbert component is given by:

qi rks r1,i rsk cos ` r2,i rsk sin (8)


Then the analytic representation is:
i rks` qi rs
kj Ai rsk e j i rks (9)
Then the modulation components are defined by:
b
Ai rks k2
i rs` qi krs2 (10)

qi rsk
k
i rs arctanp q (11)
i rsk
We use MMSA to define the amplitude component. The frequency component contains too
much noise of the reflective wave and it is not so convenient to use this component. In monogenic
representation, the local amplitude represents the local energetic information, whereas the phase
depicts the local structural information. Thus, the monogenic signal fulfills the invariant property
with respect to energy and orientation.

Figure 3 illustrates the ability of the MMSA method to extract the amplitude component. With
a fixed Gaussian window, we use a tensor-based method to estimate the orientation. Figure3a is a
GPR experiment data profile. It contains recorded point data with a mean frequency of 300 MHz
and a time window of 100 ns. The horizontal and vertical coordinates are the trace and sample
points, respectively. The number of traces is 540 and the number of sampling points for each trace
is 512. Figure 3b illustrates the three amplitude component image extraction result under three level
monogenic signal analysis. Three amplitude components are depicted in one picture. As compared
with Figure 3a, the ground surface has obvious direct waves, and the first amplitude component
suppresses the direct wave and reflective wave to a large extent. This representation not only
enhances the target reflection, but also eliminates the redundant echo. Figure 3c shows the enlarged
second amplitude-component. The amplitude component makes the target signals obvious.

30343
compared with Figure 3a, the ground surface has obvious direct waves, and the first amplitude component

suppresses the direct wave and reflective wave to a large extent. This representation not only enhances
the target
Sensors reflection, but also eliminates the redundant echo. Figure 3c shows the enlarged second
2015, 15, 3034030350
amplitude-component. The amplitude component makes the target signals obvious.

(a) (b) (c)

Sensors FigureFigure153. Example of amplitude components extracted by the MMSA method. (a) Input image;
2015, 3. Example of amplitude components extracted by the MMSA method. ( a) Input 6
(b) Wavelet-domain amplitude estimation. Three amplitude components under three level monogenic
image; ( b)analysis.
wavelet Wavelet-domain
The amplitude amplitude
componentsestimation.
eliminate the Three
direct andamplitude components
surface-wave noise; ( ) The cunder
3.2. Extraction
three level themonogenic
Region of Interest
wavelet analysis. The amplitude components eliminate the direct
enlarged second amplitude component.
and surface-wave noise; (c) The enlarged second amplitude component.
3.2.ExtractiontheRegionofInterest
Cheng proposed the Region of Interest (ROI) extraction method [10]. Based on the analysis of GPR
Cheng proposed the Region of Interest (ROI) extraction method [ 10]. Based on the analysis of A-Scan signals
statistical features, a windowed statistical method is used to extract the ROI from a GPR A-Scan
large amount
signal's
a large
of data.
statistical
amount features,
of data.a windowed statistical method is used to extract the ROI from
In detail, the variance of the B-scan data is calculated. Then the normalized variance of the data
InWedetail, the variance of the B-scan data is calculated. Then thetarget
normalized
calculated. Weisnote
is calculated. that the
note variance
that is significantly
the variance largerlarger
is significantly in the intarget
the The variance
area.area. ROI is of
The ROI isthe
extracteddata
based on
extracted
based on the normalized variance being bigger than a xed number. The rst amplitude component
the normalized variance being bigger than a fixed number. The first amplitude component extracted by the
extracted by the MMSA is shown in Figure 4a. Then the normalized variance of amplitude component
MMSA is shown in Figure 4a. Then the normalized variance of amplitude component is illustrated in
is illustrated in Figure 4b. The extraction of ROI of the amplitude is based on the normalized variance Figure 4b. The
extraction of ROI of the amplitude is based on the normalized variance with a lower with a lower
threshold. The extraction result is shown in Figure 4c. This method gets rid of the threshold. The extraction result is shown
same depth.
the targets
in Figure accurately
4c. and it gets
This method is especially
rid of theused
falsefor the targets
targets at almostfalse
and locates the targets
same depth.
and locates the targets accurately and it
is especially used for the targets at almost the

(a) (b) (c)


Figure 4. Example of the extraction of Region of Interest. (a) The first amplitude modulation component
Figure 4. Example of the extraction of Region of Interest. ( a) The first amplitude
extracted by using the MMSA; ( ) The normalized variance of amplitude component; (c) The extraction
modulation component
of ROI of the extracted
amplitude based by using variance
on the normalized bthe MMSA; (b) The
with a lower normalized variance of
threshold.
amplitude component; (c) The extraction of ROI of the amplitude based on the normalized
3.3.HoughTransform
variance with a lower threshold.
We briefly introduce the Hough transform as follows. We rearrange Equation (1):
3.3. Hough Transform c
4
t0 t2 2 px x0 q2 (12)
v
We briefly introduce the Hough transform as follows. We rearrange Equation (1):

4
t0 = t 2 - 2
( x - x0 ) 2 (12)
v
30344

where ( x, t ) represents the reflection hyperbolas coordinates and x0 , t 0 , and v are the parameters. We
Sensors 2015, 15, 3034030350

wherepxt, q represents the reflection hyperbolas coordinates and x0 , t0 , and v are the parameters.
We first limit the image area by the ROI extraction algorithm, so the computational time decreases
dramatically. The Hough transform used in our paper was shown in [ 11].
In our method, ROI limits the area before the Hough transform. In order to reduce noise and
Sensorsartifacts,
2015, 15 the image in marked regions is smoothed with Gaussian filters. Then by a canny edge 7
detector, this image is converted into a binary image. Then using the Hough transform, the position of the target
is determined. In detail we define the number of accumulator-cells for x0 and t0 using the edge pixels.
varies in a given interval assuming the range is not so big, which can be set by the user. Then
we The propagation
transform velocity
every edge
characteristics pixel
especially( xfor
, t )large
is an importantandparameter
every to obtain value
discrete
range of
the bestof
tting.
areas. The velocity vvintoIt varies
in the
for different
theHough
Hough plane
transform
soilwith Equation (1).
varies in a
given interval assuming the range is not so big, which can be set by the user. Then we transform every
Finally, we draw the location result in the original image (see Figure 5). Firstly, using the Gabor filter,
edge pixel pxt, q and every discrete value of v into the Hough plane with Equation (1). Finally, we draw
the image in the extracted region is smoothed. By a canny detector, the image becomes a binary image
the location result in the original image (see Figure 5). Firstly, using the Gabor lter, the image in the
(shown in the middle of Figure 5). Then we use Hough transform in the ROI region. The calculated
extractedmiddle
region isofsmoothed.
Figure 5).By a canny
Then detector,
we use Hough thetransform
image becomes
in theaROI
binary imageThe
region. (shown in the hyperbolas are
calculated
drawn in red color.
hyperbolas are drawn in red color.
Canny edge Hough
detector transform

Figure 5. The result of the Hough transform. The image is in the ROI region (on the left). Then the
Figure 5. The
canny edge result
red color.
detects of theisHough
result transform.
shown. Then The
the Hough image result.
transform is in the
The ROI
left). region
hyperbola (on the
is drawn in

Then the canny edge detects result is shown. Then the Hough transform result. The
hyperbola
4. is drawn in red color.
Results
We use GPR numerical simulation images and experimental images to assess the target detection.
4. Results
The electromagnetic simulator GprMax is used to generate simulation images using the
finite difference time domain (FDTD) method. RIS radar is applied to collect the experimental data.
We use GPR numerical simulation images and experimental images to assess the target detection.
4.1.SimulationResults
The electromagnetic simulator GprMax is used to generate simulation images using the finite
The whole simulation process is shown in Figure 6. Figure 6a illustrates a scenario of metal
blockstime
difference in three distinct(FDTD)
domain steps. Thismethod.
is a simulated
RIS image.
radar The width and
is applied todepth are 2.8
collect themexperimental
and 3 m, data.
respectively. The background medium is dry sand and three perfect conductor metal cylinders
4.1. Simulation
with 0.01 Results
m diameter are buried at the same depth of 0.45 m. The interval of the three metal targets
is 1 m. In the simulation model, the grid in the x and y direction is 0.01 m and detection step is
The 0.02
wholem insimulation process
the x direction. is shown
All the in Figure
boundaries 6. Figure
are Perfectly 6a illustrates
Matched a scenario
Layers (PMLs). of metal blocks
The parameter
in three distinct steps. This is a simulated image. The width and depth are 2.8 m and 3 m,
of the current definition is a Hertzian dipole with unit amplitude and center frequency of a respectively.
Ricker
wavelet of 600 MHz, and the time window for each trace is 25 ns. There are 141 traces in total
The background mediumThe
and 1060 samples. is dry sandbetween
interval and three perfect
samples conductor
is 2.3587 metal cylinders with 0.01 m diameter
1011 s. The actual electromagnetic
are buried
waveatvelocity
the same depth
is 14.8 of 0.45
cm/ns. Them. The GPR
related interval
image ofisthe threeinmetal
shown Figure targets
6b which is
is 1 m. In thebysimulation
generated
the
model, stepGprMax
the grid simulator. The extraction result of the first amplitude component
in the x and y direction is 0.01 m and detection step is 0.02 m in the of the xMMSA
direction. All
is shown in Figure6c, where the reflection signals and the noise can be suppressed and the
the boundaries are Perfectly Matched Layers (PMLs). The parameter of the current denition is a
target reflection can be enhanced after the extraction of the amplitude component. Then the ROI
Hertzian dipole with
extraction unitisamplitude
method and
used to get ridcenter
of the frequency ofand
false targets a Ricker
locate wavelet of accurately.
the targets 600 MHz, The and the time
result is illustrated in Figure6d. The results allow one to the find the three patterns correctly as can be
window for each trace is 25 ns. There are 141 traces in total and 1060 samples. The interval
seen in Figure 6e.
between samples is 2.3587 1011 s. The actual electromagnetic wave velocity is 14.8 cm/ns. The
related GPR image is shown in Figure 6b which is generated by the GprMax simulator. The
extraction result of the first amplitude component of the MMSA step is shown in Figure 6c, where the
30345
reflection signals and the noise can be suppressed and the target reflection can be enhanced after the
extraction of the amplitude component. Then the ROI extraction method is used to get rid of the false
Sensors 2015, 15, 3034030350
Sensors 2015, 15 8

Tx /Rx

h = .045m
. Target
l= 0.4m l = 1m l = 1m

l = 3m

Sand

l = .28m

(a) (b)

(c) (d)
t

Tx/Rx

Target

x=0.4m x=1.4m x=2.4m


t=0.45m t=0.45m t=0.45m

l = 3m

S and

l = .28m
x
(e) (f)
Figure 6. Example of MMSA detection result. (a) The location in reality; (b) GPR image; (c) Extraction
Figureof 6. Example of MMSA detection result. (a) The location in reality; (b) GPR image;
the amplitude component using MMSA; ( ) Extraction the ROI of the amplitude component; ( )eThe d
(c) Extraction of the
hyperbola result of theamplitude component
Hough transform; usingestimation
( ) The position MMSA; d() of
result Extraction
the detectionthe ROI of the
method.
f
amplitude component; ( e) The hyperbola result of the Hough transform; ( f) The position
Another result is presented in Figure 7. Figure 7 is an image simulated with the FDTD method.
estimation result of the detection method.
The width and depth are 2.8 m and 3 m, respectively. There are three perfect metal point targets
with the same x coordinates. The depth of the top one is 0.2 m below the ground surface and the
Another result
interval is presented
of the in Figure
other two targets 7. m.
is 0.25 Figure 7 is an receives
The antenna image simulated with2the
a trace for each cm.FDTD method.
There are The
width and
141depth aretotal
traces in 2.8which
m andcontain
3 m, respectively. There
1060 samples. The are between
interval three perfect metal
samples point targets
is 2.3587 1011 s. with
The the
same x actual electromagnetic
coordinates. waveofvelocity
The depth the topis one
14.5 is
cm/ns.
0.2 mEach detection
below
transform. This method detects the three patterns correctly.
result issurface
the ground processed bythe
and the interval
Hough of the

other two targets is 0.25 m. The antenna receives a trace for each 2 cm. There are 141 traces in total
11
which contain 1060 samples. The interval between samples is 2.3587 10 s. The actual
electromagnetic wave velocity is 14.5 cm/ns. Each detection result is processed by the Hough
transform. This method detects the three patterns correctly.

30346
Sensors 2015, 15, 3034030350
Sensors 2015, 15 9

Tx /Rx

h = 0 2.m
Target
h = 0 25.m

h = 0 25.m
3m

Sand

2 8.m
(a) (b)
t
Tx /Rx

x=1.4m
t=0.2m Target

x=1.4m
t=0.45m
3m
x=1.4m
t=0.7m
Sand

t x
2 8.m

(c) (d)

FigureFigure
7. One 7. Oneexample
more more example
of usingofMMSA.
using MMSA. ( a) Ground
(a) Ground truth of truth of the (b
the image; image;
) GPR (b ) GPR
image; (c) The
image;
hyperbola (c) The
result hyperbola
of the result of the(d)Hough
Hough transform; transform;
The position ( d) Theresult
estimation position estimation
of the result
detection method.
of the detection method.
4.2.ExperimentResults
4.2. Experiment Results
Then we get GPR images in realistic situations. The first example is shown in Figure8a, which
is data from a real GPR experiment. Four metal spheres of 0.02 m diameter were buried under dry
Then we get GPR images in realistic situations. The first example is shown in Figure 8a, which is
sand. These spheres were located side by side. The figure records the detection data with mean
frequency
Theseof
data from 200 GPR
aspheres
real MHz and time
wereexperiment.
located window
side by side.
Four ofThe
metal80 figure
ns. Finally,
spheresrecordsthe
thedetection
of 0.02 detection
m diameter result
data is attached
with
were to thedry
mean frequency
buried under picture
of sand.
(see Figure 8b). Further, x0 and v are estimated by fitting the extracted hyperbola. After the Hough
200 MHz and time window of 80 ns. Finally, the detection result is attached to the picture (see
transform,
Figure we 8b).get that thex velocity of the data is v 1.38 108m/s, which is tested with the position
Further, 0 and v are estimated by fitting the extracted hyperbola. After the Hough
of the targets in the real situation. As illustrated in Figure8c, there are four targets in the ground.
transform, we get that the velocity of the data is v = 1.38 10 8m/s, which is tested with the position of
According to the extracted hyperbolic vertex coordinates, the antenna position of the target is shown in
the targets
Figure 8c. Forinexample,
the real thesituation.
vertex As illustrated
coordinate in Figure
of the 8c, isx
left target there are
50, four
so thetargets in position
antenna the ground.
of the
1
targetAccording to theisextracted
is 3 m, which the same hyperbolic vertex coordinates,
as the position theinantenna
of the target reality.position
Besides,of the
thetarget
vertexis coordinates
shown in 1
Figure 8c. For example, the vertex coordinate of the left target is x = 50 , so the antenna position of
axis of the hyperbola represents the depth of the target. The vertex coordinate of the left target is
t1 target
the 97. is 3 m, which is the same as the position of the target in reality. Besides, the vertex
coordinates axis of the hyperbola represents the depth of the target. The vertex coordinate of the left
A second
target is t1 example provided by the MMSA method is shown in Figure 9. Three similar landmines
= 97 .
with a plastic case were buried under dry sand. The mines are Type-72A. The diameter of the
landmine is 7.7 cm and the height is 3.7 cm. Figure 9a shows the final detection result with the proposed MMSA
method. Note that the proposed detection scheme provides a high detection rate even though the
target is not so obvious. Figure9a is a real GPR data profile. It recorded the point data with a mean
frequency of 200 MHz and time window of 50 ns. The detection result is shown in Figure9b,
which illustrates that all patterns are identified with high precision.

The last example is shown in Figure10, which illustrates the effects of the proposed MMSA
detector and the final detection results. Note that the proposed detection scheme provides a high
detection rate. Figure 10a is a real GPR data profile. It is recorded landmine detection data with a mean

30347
(a)
Sensors 2015, 15, 3034030350 (b) (c)

Figure 8. Example of test metal point targets data. ( a) GPR image; ( b) Processed result
frequency of 200 MHz and time window of 100 ns. This is a two dimensional image. The detection
with is
result detected patterns. The detected hyperbola is the red color line; ( c) The position
illustrated
Sensors 2015, 15 in Figure10b, which shows that all targets are localized. 10
estimation results of the real point data.

A second example provided by the MMSA method is shown in Figure 9. Three similar landmines
with a plastic case were buried under dry sand. The mines are Type-72A. The diameter of the
landmine is 7.7 cm and the height is 3.7 cm. Figure 9a shows the final detection result with the
proposed MMSA method. Note that the proposed detection scheme provides a high detection rate
even though the target is not so obvious. Figure 9a is a real GPR data profile. It recorded the
point data with a mean frequency of 200 MHz and time window of 50 ns. The detection result is
shown in Figure 9b, which illustrates that all patterns are identified with high precision.

The last example is shown in Figure 10, which illustrates the effects of the proposed MMSA
detector and the final detection
( a) results. Note that
(b) the proposed detection(c) scheme provides a high
detection rate. Figure 10a is a real GPR data profile. It is recorded landmine detection data with a mean
Figure
Figure 8. Example
8. Example of testofmetal
test metal
point point targets
targets data. data. ( image;
(a) GPR a) GPR(b)image; ( b) result
Processed Processed
with result
detected
frequency ofwith
200 detected
MHz and time The
patterns. window of 100
detected ns. This
hyperbola is ared
c the
is two dimensional
color line; ( c) The image. The
position detection
pointpatterns.
data. The detected hyperbola is the red color line; ( ) The position estimation results of the real
result is illustrated in Figure 10b, which shows that all targets are localized.
estimation results of the real point data.
A second example provided by the MMSA method is shown in Figure 9. Three similar landmines
with a plastic case were buried under dry sand. The mines are Type-72A. The diameter of the
landmine is 7.7 cm and the height is 3.7 cm. Figure 9a shows the final detection result with the
proposed MMSA method. Note that the proposed detection scheme provides a high detection rate
even though the target is not so obvious. Figure 9a is a real GPR data profile. It recorded the
point data with a mean frequency of 200 MHz and time window of 50 ns. The detection result is
shown in Figure 9b, which illustrates that all patterns are identified with high precision.

The last example is shown in Figure 10, which illustrates the effects of the proposed MMSA
detector and the final detection results. Note that the proposed detection scheme provides a high
detection rate. Figure 10a is a real GPR data profile. It is recorded landmine detection data with a mean
frequency of 200 MHz and time window of 100 ns. This is a two dimensional image. The detection
(a)which shows that all targets are localized.
result is illustrated in Figure 10b, (b)
Figure 9. Example of landmine detection results. (a) GPR image; ( ) Processed image withb
Figure 9. Example of landmine detection results. (a) GPR image; (b) Processed image
Sensors 2015, 15 patterns.
detected 11
with detected patterns.

(a) (b)

Figure 9. Example of landmine detection results. (a) GPR image; (b) Processed image
with detected patterns.
(a) (b)
Figure 10. Example of test metal point targets data. (a) Real GPR image; (b) Processed results with
Figure 10. Example of test metal point targets data. ( a) Real GPR image; ( b) Processed
detected patterns.
results with detected patterns.

We compare different aspects of the output30348


of K-means+CEM [12], Hyperbolas Automatic
Detection algorithm (HADA) [13] with the MMSA method:
1. How many hyperbolas are recognized correctly (true positives) and how many non-hyperbolas/
Sensors 2015, 15, 3034030350

We compare different aspects of the output of K-means+CEM 12],


[ Hyperbolas Automatic
Detection algorithm (HADA) [13] with the MMSA method:

1. How many hyperbolas are recognized correctly (true positives) and how many non-hyperbolas/
non-targets are recognized as hyperbolas (false positives)?
2. The distance error of all correctly recognized targets.

For the first aspect we investigate three different measurements:


number of true positives
Recall (13)
number of all actual hyperbolas

number of true positives


Precision (14)
num. of all hyperbolas(true+false positives)
PrecisionRecall

F measure 2 (15)
Precision `
Recal
Table 1 show the different results of MMSA, HADA and K-means+CEM. The comparison in Table1
shows that MMSA improves Precision and Recall as well as the position estimates. These improvements
result from the fact that MMSA detects every hyperbola cluster. For K-means+CEM many position
estimates have two errors larger than 10 cm, resulting in bad values for Recall and Precision. For
correctly recognized targets there is still a distance error of a few centimeters. The reason is that
K-means+CEM method has problems with nearby and tangent hyperbolas as well as with background
noise. For the HADA algorithm, the values for Recall and Precision are higher than with K-means+CEM,
but there are still two position estimate errors larger than 10 cm. MMSA achieves much better results than
the other methods with the highest values of Recall and Precision.

Table 1. Results for data set 1 (RecallPrecision


, and F-measure with respect to hyperbolas).

Method Recall (Hyperbolas) Precision (Hyperbolas) F-measure (Hyperbolas) Distance Error(cm)


K-means+CEM 0.62 0.62 0.62 34.8
HADA 0.75 0.75 0.75 22.8
MMSA 1 1 0 5.8

5. Conclusions
In this essay, target detection using MMSA has been demonstrated. This method detects buried
objects and estimates their position correctly. The method combines the characteristics of wavelet
and energy extraction. The amplitude component of the data is extracted using an unsupervised
strategy based on MMSA. After the extraction of the amplitude component of the original date,
background noise and direct wave are removed to large extent. We use the ROI extraction method to
locate the genuine target area among spurious reflections. Then the labeled area is processed by the
Hough transform to search the targets parameters. Finally, we estimate the horizontal and vertical
position of the target. Experiments on many different GPR images show that the proposed method
provides a general hyperbola-detection system with good results.

The authors thank K. Balac (KB) for her implementation of the Laplace/Riesz transform as a
Acknowledgments:
JAVA Plug in for Image and Kaifu Niu for his work on the language editing. This research is financially supported by the
National Natural Science Foundations of China (under Grant Nos. 11201120, 61201389, 61201390); Key Scientific Research
Project Program of Universities of Henan Province (under Grant No. 15A120001); Plan For Scientific Innovation Talent of
Henan University of Technology(under Grant No. 2012CXRC07); Provincial Colleges and Universities on Basic Research
Fund of Henan Technology University(under Grant No. 2014YWQQ11) and China Grain Public Service Research Projects
(under Grant No. 201413003). Open fund of Key Laboratory of Grain Information Processing and Control (under Grant No.
KFJJ-2015-102).

Qiao, the principal author, contributed mostly to the literature survey and digestion in
Author Contributions:
the classicaland new fields of information theory, radar and MMSA. She was also responsible for the research

30349
Sensors 2015, 15, 3034030350

design, results analysis and interpretation and writing the paper. The primary contribution of Qin, the co-author,
include the literature review, research-related experimentation and manuscript formatting.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Delbo, S.; Gamba, P.; Roccato, D. A fuzzy shell clustering approach to recognize hyperbolic signatures
in subsurface radar images.
IEEETrans.Geosci.RemoteSens. 2000, 38, 14471451. [CrossRef]
2. Gamba, P.; Lossani, S. Neural detection of pipe signatures in ground penetrating radar images. IEEETrans.
Geosci.RemoteSens. 2000, 38, 790797. [CrossRef]
3. Kaneko, T. Radar image processing for locating underground linear objects. IEICETrans. 1991, 74, 34513458.
4. Shihab, S.; Al-Nuaimy, W.; Eriksen, A. Image processing and neural network techniques for automatic
detection and interpretation of ground penetrating radar data. In Proceedings of the 6th International
Multi-Conference on Circuits, Systems, Communications and Computers, Cancun, Mexico, 1216 May
2002. Capineri, L.; Grande, P.; Temple, J.A.G. Advanced image-processing technique for real-time
5. interpretation of ground-penetrating radar images.
Int.J.ImagingSyst.Technol. 1998, 9, 5159. [CrossRef]
6. Unser, M.; Sage, D.; van de Ville, D. Multiresolution monogenic signal analysis using the Riesz-Laplace
wavelet transform. IEEETrans.ImageProcess. 2009, 18, 24022418. [CrossRef] [PubMed]
7. Dong, G.; Wang, N.; Kuang, G. Sparse representation of monogenic signal: With application to target
recognition in SAR images. IEEESignalProcess.Lett. 2014, 21, 952956.
8. Dong, G.; Kuang, G. Target recognition in SAR images via classification on Riemannian manifolds.
IEEEGeosci.RemoteSens.Lett. 2015, 12, 199203. [CrossRef]
9. Felsberg, M.; Sommer, G. The monogenic scale-space: A unifying approach to phase-based image processing
in scale-space. J.Math.ImagingVis. 2004, 21, 526. [CrossRef]
10. Chen, D.L. ROI Extraction Research of Ground Penetrating Radar Data. Bachelor Thesis, National University
of Defense Technology, Changsha, China, 2003.
11. Huang, K.; Fu, K.; Sheen, T.; Cheng, S. Image processing of seismograms: Hough transformation for the
detection of seismic patterns; thinning processing in the seismogram. PatternRecog. 1985, 18, 429440.
[CrossRef]
12. Chen, H.; Cohn, A.G. Probabilistic robust hyperbola mixture model for interpreting ground penetrating
radar data. In Proceedings of the 2010 International Joint Conference on Neural Networks, Barcelona, Spain,
1823 July 2010.
13. Simi, A.; Bracciali, S.; Manacorda, G. Hough transform based automatic pipe detection for array GPR:
Algorithm development and on-site tests. In Proceedings of the 2008 IEEE Radar Conference, Rome,
Italy, 2630 May 2008; pp. 16.

2015 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons by
Attribution (CC-BY) license (http://creativecommons.org/licenses/by/4.0/).

30350