In this paper, we present a method to detect the position of a 3D object in scattering media by using the axially distributed sensing (ADS) method. Due to the scattering noise of the elemental images recorded by the ADS method, we apply a statistical image processing algorithm where the scattering elemental images are converted into scatterreduced ones. With the scatterreduced elemental images, we reconstruct the 3D images using the digital reconstruction algorithm based on ray backprojection. The reconstructed images are used for the position detection of a 3D object in the scattering medium. We perform the preliminary experiments and present experimental results.
I. INTRODUCTION
Visualization and detection of a 3D object in optically scattering media is a challenging problem. An improved visualization method of scattered images can be applied to various applications such as imaging in fog, underwater, security, and medical diagnostics
[1

7]
. Some 3D multiperspective approaches including integral imaging and axially distributed sensing (ADS) have been studied for visualizing objects in the scattering media
[8

9]
. Among them, the ADS scheme can provide simple recording architecture by translating a camera along its optical axis. The recorded highresolution elemental images generate the clear 3D plane images for the partially occluded 3D objects.
In this paper, we present a position detection method for a 3D object in scattering media by using the ADS method
[10

12]
. In the proposed method, we introduce the statistical signal processing to reduce the scattering effect in elemental images and the position detection process with nonlinear correlation operation is newly applied to the scattered images in ADS. The proposed method is divided into four parts: (1) pickup (2) statistical image processing (3) digital reconstruction and (4) position detection. In the pickup part using the ADS method, the camera is moved along a common optical axis and multiple longitudinal 2D images of the scene are recorded. The recorded multiple 2D images of the 3D object in the scattering media are referred to as scattering elemental images. In the second statistical processing part, to reduce the scattering effects of the recorded elemental images, a statistical image processing technique is applied to the scattering elemental images. Then, the statistically processed elemental images are used to reconstruct the 3D plane images using the digital reconstruction algorithm based on ray backprojection in the third part. In the final part, the reconstructed plane images are used for 3D object detection. To show the feasibility of the proposed method, we carry out the preliminary experiments for a scattering 3D object behind a diffuser screen as the scattering medium.
II. RECOGNITION METHOD USING ADS
 2.1. Pickup Method in ADS
The pickup part of the ADS method is shown in
Fig. 1
. We translate a single camera along the optical axis and record multiple 2D images of the scene observed through the scattering media. We suppose that the object is located at a distance
z_{0}
from the 1st camera position. The distance between the cameras is Δz. We record
K
elemental images by moving the camera. The first camera is
k
=1 at
z_{k}
=
z_{0}
and the camera closest to the object becomes by
k
=
K
at
z_{k}
=
z_{0}
(
K
1)
Δz
. Due to the different distance from the object for each camera, the object is recorded in each elemental image with a different magnification. Each magnification ratio is calculated as
M_{k}
=
z_{k} /g
where
g
is the focal length of the camera.
Pickup part of ADS.
 2.2. Statistical Processing of Elemental Images
When a 3D object is located in scattering media, the recorded image may contain severe scattering noise which can prevent the correct visualization and detection for the 3D object. In this paper, therefore, we wish to reduce the noise by using statistical image processing. To do so, we apply statistical image processing techniques to the scattering elemental images. Let us assume that degradation caused by the diffuser can be modeled as Gaussian
[6]
. The statistical image processing procedure to reduce the scattering effect in the elemental images is shown in
Fig. 2
. First, we estimate the unknown parameter
μ
(mean) of the
Gaussian
distribution using the
Maximum Likelihood Estimation
(MLE). Here we consider that the scattering degradation function is composed of many superimposed Gaussian random variables with local area (
w_{x}
×
w_{y}
) of an elemental image with (
N_{x}
×
N_{y}
) pixels. Let us denote these random variables as
X_{ij}
(
m
,
n
). Then, MLE is calculated by
Statistical image processing for the recorded elemental images.
where
L
(··) is the likelihood function,
i
=1, 2, ⋯,
N_{x}
−
w_{x}
+1,
j
=1, 2, ⋯,
N_{y}
−
w_{y}
+1,
m
=1, 2, ⋯,
w_{x}
, and
n
=1, 2, ⋯,
w_{y}
. And, the mean parameter
of the
Gaussian
distribution is estimated as
The estimated elemental images are obtained by subtracting
. The estimated elemental images are given by
This estimation process enables us to reduce the scattering noise in the elemental images.
Gamma
(
γ
) correction is then used for the estimated elemental images. Using
Gamma
correction, we can manipulate the histogram to improve the contrast ratio
[8]
. Gamma correction is a nonlinear method to control the overall luminance of a still image. Gamma correction is typically used to adjust still images for accurate reproduction on physical displays. In this case, it is used to compensate for the nonlinear relationship between pixel values and the intensity being displayed. Therefore, for the estimated elemental images, it may be expressed by the following powerlaw relationship:
where
S_{uncorrected}
and
S_{corrected}
are the uncorrected and corrected elemental images respectively, and
γ
is the gamma value.
Next, we apply histogram equalization and matching to the corrected elemental images. This process can remove artificial gray levels from the elemental images. We first calculate a cumulative distribution function (CDF) and this is given by
where
h
,
h_{e}
, and
q
are continuous gray levels of the histogram stretched image, the histogram equalized image, and histogram matched image.
p_{h}
(
w
) and
p_{q}
(
t
) are the continuous probability density functions corresponding to the h’s and q, and
w
and
t
are the integration variables. Equations (5) and (6) should be followed by
G
(
q
) =
T
(
h
) where the inverse of the
G
operation yields the restored image. Finally, we can obtain the new scatterreduced elemental images.
 2.3. Digital Reconstruction in ADS
In the proposed method, the third part is the digital reconstruction using the scatterreduced elemental images. This is shown in
Fig. 3
. It is based on ray backpropagation through virtual pinholes. This reconstruction algorithm numerically implements the reverse of the pickup process. In other words, all scatterreduced elemental images are back projected and magnified by different magnifications
M_{k}
(
z
) if the reconstruction plane is located at
z
. The magnified scatterreduced elemental images are superimposed at the same reconstruction plane. Finally, we can obtain the reconstructed plane image at distance
z
. Then, the reconstructed plane image
r
(
x, y, z
) is the summation of all the magnified elemental images
S_{k}
and is given by
Ray diagram for digital reconstruction.
For faster calculation, the reconstruction process may be normalized by the magnification ratio
M_{1}
. This becomes
 2.4. Position Detection Process
In the last part of the proposed method, the plane image
r
(
x, y, z
) reconstructed at
z
distance from the digital reconstruction part is used for 3D object position detection. When a plane image of the target object is reconstructed at
z
distance, correlation is performed with the reference image
f
to detect the position of the 3D object. The reference image can become a clear object image without scattering medium. Then, we simply obtain the correlation peak result from nonlinear correlation operation as given by
where
α
is the nonlinear parameter and means the strength of the applied nonlinearity. In general, the choice of
α
depends on the type of noise and object
[14]
. It could vary within the range [0, 1].
F
and
R
are the Fourier transforms of
f
and
r
, respectively. And
ϕ_{F}
and
ϕ_{R}
represent the phase angle of
f
and
r
, respectively. In Eq. (9), the correlation peak indicates the position detection performance between the reference image and the reconstructed plane image.
III. EXPERIMENTS AND RESULTS
To demonstrate the proposed position detection method for a 3D object in scattering media, we carried out optical experiments. The experimental setup is shown in
Fig. 4
where a ‘car’ object is positioned approximately 440 mm away from the first camera position. We placed a diffuser at approximately 50 mm in front of the object. The thickness of the diffuser was 1 mm. We used a camera with 2184×1456 pixels with 8.2 um pixel pitch. An imaging lens with focal length
g
=70 mm was used. We used the camera with the smallest aperture to obtain the maximum depth of field. Then, the pinhole gap (
g
) becomes 70 mm in the digital reconstructions as shown in
Fig. 3
. The camera is moved with a step of 5 mm and a total of
K
=41 elemental images are recorded within a total displacement distance of 200 mm. The first elemental image and the 41st elemental image are shown in
Fig. 4(c)
and
4(d)
, respectively. We can see that the recorded elemental images were degraded by the scattering noise.
(a) Original object (b) Experimental setup. (c) 1st scattering elemental image (d) 41st scattering elemental image.
To reduce the scattering noise in the elemental images, the statistical image processing technique was applied to them as described in
Fig. 3
.
Figure 5
shows the example of the 21st elemental image produced by the statistical image processing technique. The 21st elemental image shown in
Fig. 5(a)
was scatterreduced as shown in
Fig. 5(b)
. The statistical image processing was repeated to all the recorded elemental images. Then, we can obtain the 41 scatterreduced elemental images.
(a) ‘car’ image within 21st elemental image (b) Statistically processed images.
To reconstruct the 3D images, the scatterreduced 41 elemental images were computationally back projected through virtual pinholes according to Eq. (8). The 3D plane images were obtained according to the reconstruction distance. Some reconstructed plane images are shown in
Fig. 6(a)
for three different reconstruction distances (340 mm, 440 mm, and 540 mm). When the distance of reconstruction plane was 440 mm, where the ‘car’ object is originally located, the reconstructed image is well focused after reducing the scattering noises through the statistical image processing. For comparison, the plane images are shown in
Fig. 6(b)
when using the conventional method without the statistical image processing.
Reconstructed plane images at different distances (a) Proposed method (b) Conventional method.
To detect the position of the 3D object, the nonlinear correlation operation as described in Eq. (9) was used for the original reference image and the reconstructed plane images. The correlation peak results are shown in
Fig. 7
according to the reconstruction distances from 340 mm to 540 mm. Here the nonlinear parameter was 0.6. We can see that the correlation peaks are very low in the conventional method. This is because all of the reconstructed plane images have heavily scattered noise regardless of the reconstruction distance. On the other hand, we can see the highest correlation peak at 440 mm where the object was originally located in the result of the proposed method. At the original position of the object, the 2D correlation output was shown in
Fig. 8(a)
when the original image and our scatterreduced plane image were used. For comparison, the correlation output in the conventional method is very low when the scattered plane image was used as shown in
Fig. 8(b)
.
Figure 9
shows the graph of the maximum correlation peaks according to the nonlinear parameter (
α
) in Eq. (9). These results reveal that the scatterreduced plane image provides the higher position detection performance compared with the scattered plane image for all
α
values.
Correlation peak results according to the reconstruction distance.
Nonlinear correlation results when α=0.6 (a) Proposed method (b) Conventional method. x and y mean the pixel indexes of the correlation output.
Nonlinear correlation results according to the α value.
IV. CONCLUSION
In conclusion, the ADSbased position detection method for a 3D object in scattering media was presented. The elemental images recorded by the ADS method contain severe scattering noise. To reduce the scattering noise, we applied a statistical image processing algorithm to the scattering elemental images. Thus, we were able to obtain the improved 3D plane images using the digital reconstruction algorithm based on ray backprojection. Using them, we compare the correlation performance for the proposed method and the conventional method. The experimental results reveal that the proposed method is superior to the conventional method in terms of correlation operation.
Color versions of one or more of the figures in this paper are available online.
Tyo J. S.
,
Rowe M. P.
,
Pugh E. N., Jr.
,
Engheta N.
1996
Target detection in optically scattering media by polarizationdifference imaging
Appl. Opt.
http://dx.doi.org/10.1364/AO.35.001855
35
1855 
1870
Narasimhan S. G.
,
Nayar S. K.
2002
Vision and the atmosphere
Int’l J. Computer Vision
http://dx.doi.org/10.1023/A:1016328200723
48
233 
254
Bartolini L.
,
De Dominicis L.
,
Ferri de Collibus M.
,
Fornetti G.
,
Guarneri M.
,
Paglia E.
,
Poggi C.
,
Ricci R.
2005
Underwater threedimensional imaging with an amplitudemodulated laser radar at a 405 nm wavelength
Appl. Opt.
http://dx.doi.org/10.1364/AO.44.007130
44
7130 
7135
Treibitz T.
,
Schechner Y. Y.
2009
Active polarization descattering
IEEE Trans. PAMI
http://dx.doi.org/10.1109/TPAMI.2008.85
31
385 
399
Gilbert G. D.
,
Pernicka J. C.
1967
Improvement of underwater visibility by reduction of backscatter with a circular polarization technique
Appl. Opt.
http://dx.doi.org/10.1364/AO.6.000741
6
741 
746
Cho M.
,
Daneshpanah M.
,
Moon I.
,
Javidi B.
2011
Threedimensional optical sensing and visualization using integral imaging
Proc. IEEE
http://dx.doi.org/10.1109/JPROC.2010.2090114
99
556 
575
Moon I.
,
Javidi B.
2008
Threedimensional visualization of objects in scattering medium by use of computational integral imaging
Opt. Express
http://dx.doi.org/10.1364/OE.16.013080
16
13080 
13089
Cho M.
,
Javidi B.
2010
Threedimensional visualization of objects in turbid water using integral imaging
J. Display Technol.
http://dx.doi.org/10.1109/JDT.2010.2066546
6
544 
547
Shin D.
,
Javidi B.
2011
3D visualization of partially occluded objects using axially distributed sensing
J. Display Technol.
http://dx.doi.org/10.1109/JDT.2011.2124441
7
223 
225
Schulein R.
,
DaneshPanah M.
,
Javidi B.
2009
3D imaging with axially distributed sensing
Opt. Lett.
http://dx.doi.org/10.1364/OL.34.002012
34
2012 
2014
Shin D.
,
Javidi B.
2012
Threedimensional imaging and visualization of partially occluded objects using axially distributed stereo image sensing
Opt. Lett.
http://dx.doi.org/10.1364/OL.37.001394
37
1394 
1396
Piao Y.
,
Zhang M.
,
Shin D.
,
Yoo H.
2013
Threedimensional imaging and visualization using offaxially distributed image sensing
Opt. Lett.
http://dx.doi.org/10.1364/OL.38.003162
38
3162 
3164
Cho M.
,
Shin D.
2013
3D integral imaging display using axially recorded multiple images
J. Opt. Soc. Korea
http://dx.doi.org/10.3807/JOSK.2013.17.5.410
17
410 
414
Javidi B.
1989
Nonlinear joint power spectrum based optical correlation
Appl. Opt.
http://dx.doi.org/10.1364/AO.28.002358
28
2358 
2367