In this paper we examine and compare the computational speeds of threedimensional (3D) object recognition by use of digital holography based on central unit processing (CPU) and graphic processing unit (GPU)computing. The holographic fringe pattern of a 3D object is obtained using an inline interferometry setup.The Fourier matched filters are applied to the complex image reconstructed from the holographic fringe pattern using a GPU chip for realtime 3D object recognition. It is shown that the computational speed of the 3D object recognition using GPU computing is significantly faster than that of the CPU computing.To the best of our knowledge, this is the first report on comparisons of the calculation time of the 3D object recognition based on the digital holography with CPU vs GPU computing.
I. INTRODUCTION
Threedimensional (3D) optical imaging using digital holography
[1

4]
and its applications have been investigated for 3D sensing, imaging, display, and recognition of objects
[5

15]
. This technique records a twodimensional (2D)holographic fringe pattern of a 3D object using an optical inline interferometry setup. The numerical Fresnel back propagation algorithms are applied to the holographic fringe pattern for 3D image reconstruction of the object
[2
,
4]
.Recently, a digital holographic imaging system has been explored for 3D macro or microobject recognition
[13

18]
.A variety of techniques, including matched filter, neural network and statistical sampling & inference have been applied for 3D recognition of macro or microobjects using CPU computing
[13

14
,
18]
. However, the high computational time of the digital holographic imaging system using recent CPUs seems to make it difficult to reconstruct or recognize the 3D objects in realtime. To overcome this problem,realtime holographic imaging by using a graphic processing unit (GPU) chip has been studied
[19]
.
In this paper, we propose a realtime 3D object recognition system based on digital holography using GPU computing.For fast 3D sensing and imaging of objects, a holographic fringe pattern of the object is obtained by using a singlestep or twostep inline optical interferometry setup
[14]
.The complex image of the 3D object is numerically reconstructed from the holographic fringe pattern by Fresnel back propagation using the GPU chip. For realtime 3D recognition, the Fourier matched filter
[13

14]
is applied to the reconstructed complex image using the GPU computing.For the 3D object recognition without the complex image reconstruction, the matched filter can be directly applied to the holographic fringe pattern
[13]
using the GPU computing.We examine and compare the computational speed as a performance metric for the 3D recognition systems based on digital holography using the GPU and CPU computing for two different sensing scenarios. For the first scenario we examine the computational speed of the 3D object recognition by use of digital holography based on GPU computing with the correlation peak values calculated between the reference and unknown input complex images reconstructed with the varied holographic fringe pattern sizes.For the second case of 3D object recognition without complex image reconstruction, we examine the computational speed and performance of the 3D object recognition by
The procedure of parallel implementation on GPU for realtime 3D object recognition.
use of digital holography based on GPU computing, with the correlation peak values directly computed between the reference and unknown input holographic fringe patterns with varied holographic fringe pattern sizes. We quantitatively illustrate that the calculation speed of the proposed GPU computingbased 3D object recognition can be significantly faster than that of a CPU computingbased one. Our proposed methodology for parallel implementation on GPU for realtime 3D object recognition is described in
Fig. 1
.
This paper is organized as follows. In Section 2, we present the description for digital holography used in the experiments. In Section 3, we describe the matched filter for 3D object recognition. In Section 4, we illustrate the hardware and software configurations for GPU computing.Then, in Section 5, we show the experimental results and performance of GPU computing. Finally, we conclude in Section 6.
II. FAST COMPLEX IMAGE RECONSTRUCTION OF A 3D OBJECT
In this section, the digital holography setup for realtime 3D sensing and imaging is described.
Figure 2
shows the schematic setup of phaseshifting (twostep) or singleexposure(singlestep) inline digital holography
[14]
. For the twostep phaseshifting digital holography, the interference pattern between a Fresnel diffraction pattern under a 3D object and a parallel reference beam is recorded at a CCD camera using an inline interferometry setup as shown in
Fig. 2
.
Let the complex amplitude distribution of a 3D object
Schematic diagram of twostep phaseshifting orsingleexposure inline digital holography.
U(x, y)
at the CCD plane be given as:
where
A_{U} (x, y)
and
φ_{U} (x, y)
are the amplitude and the phase of the diffracted object beam at the CCD plane,respectively. Let the parallel reference beam at the CCD plane be given as:
where
A_{R}
and
φ_{R}
are the amplitude and the phase of the reference beam, respectively, and
θ
is a phaseshifting value set to either 0 or π/2. Then, the interference pattern recorded at the CCD plane is given as:
In twostep phase shifting digital holography, the holographic fringe pattern
DH_{p}(x, y)
of the 3D object is obtained from the two interference patterns with the phase difference of π/2 and the two DC term, ∣
A_{U}(x, y)
∣
^{2}
and ∣
A_{R}
∣
^{2}
as follows
[14]
:
It can be assumed that the reference beam intensity ∣
A_{R}
∣
^{2}
can be obtained with one time measurement and the diffracted object beam intensity ∣
A_{U}(x, y)
∣
^{2}
, which depends on the 3D object, can be numerically measured with the averaging method
[14]
:
For a computational complex image reconstruction of the original 3D object, the following Fresnel diffraction is applied to the holographic fringe pattern of Eq. (4)
[2
,
13]
:
where
O(x_{0}, y_{0})
is the reconstructed complex image for the original 3D object and
z
is the reconstruction distance between the reconstruction plane and the CCD plane. It is noted that the in Eq. (5), which is multiplication of the amplitude and the phase terms, means the wavefront of the
O(x_{0}, y_{0})
original 3D object.
For fast 3D sensing and imaging of objects, a singleexposure inline digital holography (SIDH) may replace the twostep phaseshifting digital holography, which requires two different interference patterns. In the SIDH, the holographic fringe pattern
DH_{s}(x, y)
of the 3D object is obtained from the only one interference pattern as follows
[14]
:
Similarly, the reference beam intensity ∣
A_{R}
∣
^{2}
can be obtained with one time measurement and the diffracted object beam intensity ∣
A_{U}(x, y)
∣
^{2}
can be numerically measured with the averaging method
[14]
. The complex image of the original 3D object is numerically reconstructed from the holographic fringe pattern
DH_{s}(x, y)
using the Fresnel diffraction of Eq. (5). It should be noted that the reconstructed complex image contains a conjugate image which degrades the quality of the reconstructed image. However, it should be considered that the intrinsic defocused conjugate image also contains 3D information of the object.
For the computational implementation of Eq. (5), the discrete Fresnel diffraction can be given as
[13]
:
where
(k_{0}, l_{0})
and
(k, l)
are discrete spatial coordinates in the reconstruction plane and the CCD plane, respectively,(N
_{k}
, N
_{l}
) are the numbers of pixels of the CCD camera in the x and y directions and (
Δx_{0}, Δy_{0}
) and (
x_{0}, y_{0}
) are the resolution of the CCD camera in the x and y directions at the reconstruction plane and CCD plane, respectively. It is noted that the holographic fringe pattern
DH(k, l)
can be either phaseshifting holographic fringe pattern
DH_{p}(k, l)
or singleexposure inline holographic fringe pattern
DH_{s}(k, l)
.Let the above Fresnel diffraction equation be represented by
[13]
:
where
a(k_{0}, l_{0})
is
,
f(k, l )
=
and
DFT
denotes a discrete Fourier transforms. Therefore, the discrete Fresnel diffraction can be calculated by using the FFT algorithm on the computer.We can call the FFT function in Jacket Library if we need FFT operations to be directly passed to the GPU to process in parallel. To compare the computation time of the complex image reconstruction process on the CPU in sequential and the GPU in parallel, we evaluate computation time of the term
a(k_{0}, l_{0})
and
DFT
[
f(k, l)
] in Eq. (8).
III. FAST COMPLEX IMAGE RECOGNITON
The 3D pattern recognition of unknown objects is performed by evaluation of the reference complex image
Ŏ_{R}(k, l)
and unknown input complex image
Ŏ_{I}(k, l)
reconstructed from the region under consideration from the corresponding phaseshifting or single exposure inline holographic fringe pattern. In order to test digital holography using the GPU computing for realtime 3D object recognition, a discrete Fourier matched spatial filter is applied to the complex image as follows
[13]
:
where
DFT
denotes a discrete Fourier transformation and
DFT^{1}
denotes an inverse discrete Fourier transformation.The numerical calculation for the discrete matched filter can be conducted by use of a FFT algorithm. As mentioned above, the FFT and IFFT operations are implemented by calling relevant functions in Jacket Library if we need to pass them directly to the GPU to process. For 3D recognition,we measure the autocorrelation of the reference complex image and the crosscorrelation of the reference and the corresponding unknown input complex images. Then, we examine and compare the computation time of calculation of the correlations on the CPU with sequential computing and the GPU with parallel computing.
For the 3D object recognition without complex image reconstruction, the discrete Fourier matched filter can be directly applied to the holographic fringe pattern using the GPU computing. In this case, the discrete Fourier matched filter is
[13]
:
where
DH_{R}(k, l)
and
DH_{I}(k, l)
denote the phaseshifting or the singleexposure inline holographic fringe pattern for reference and unknown input, respectively. Similarly, we examine and compare the computational speed and performance of the 3D recognition on the CPU with sequential computing and the GPU with parallel computing using the phaseshifting or the singleexposure inline holographic fringe pattern.
IV. HARDWARE AND SOFTWARE CONFIGURATIONS
The CPU chip used in our experiment is Intel Pentium D945 of 3.4 GHz and the memory size is 3 Gbyes. The GPU chip used in our experiment is Geforce GTX480 which is mounted on the GPU board. The GPU specifications are processor clock of 1.4 GHz, memory clock of 1.8 GHz,480 stream processors and a memory of 1.5 Gbytes.
All source codes are written under Matlab2010a and parallel computations on the GPU are implemented by calling Jacket Library of Accerelereyes Corporation. Jacket is a powerful plugin connecting Matlab to the GPU which brings together the best of three important computational worlds;computational speed, visualization, and userfriendliness of M programming.
V. EXPERIMENTAL RESULTS
Experiments to examine the computational speed of graphic processing unit (GPU)based digital holographic imaging system for realtime 3D object recognition are presented. In the experiments, two toy cars, car I and car II were used as a reference object and as an unknown input (false class) object, respectively. We recorded the interference patterns of each 3D car at a CCD camera with 2048 × 3072 pixels and pixel size of 12 ㎛×12 ㎛ by changing the phase of the reference beam to 0 and π/2 using an inline phase shifting interferometry setup as shown in
Fig. 2
. The phaseshifting holographic fringe patterns of toy car I and car II were obtained by using Eq. (4), where we generated eight phaseshifting holographic fringe patterns with different sizes for each car, from 256×256 to 1152 × 1152 in order to see the influence of holographic fringe pattern size on the computational speed of 3D recognition with CPU or GPU computing. Each different holographic fringe pattern size was generated by extraction of the corresponding different windows in the original holographic fringe pattern with size of 2048 × 3072 pixels. Figure 3 show the real and imaginary parts of the phaseshifting holographic fringe pattern with size of 1024 × 1024 pixels for car I and car II, respectively.
For 3D imaging, the complex images of car I and car II were numerically reconstructed from the corresponding phaseshifting holographic fringe patterns at a distance of 880mm by discrete Fresnel transform in Eq. (8) using CPU or GPU computing.
Figure 4
shows the reconstructed complex images of car I and car II from the corresponding phaseshifting holographic fringe patterns, respectively.
For 3D recognition, the matched filter in Eq. (9) was applied to the reconstructed complex images of car I and the car II using CPU or GPU computing, respectively. In this experiment, the complex image size was varied from
The phaseshifting holographic fringe patterns of the toy cars with size of 1024×1024 pixels. (a) Real part of the fringe patternfor car I (b) imaginary part of the fringe pattern for car I (c) real part of the fringe pattern for car II and (d) imaginary part of the fringepattern for car II.
(a)(b) The reconstructed complex images of toy cars at 880mm from phaseshifting holographic fringe pattern with size of1024×1024 pixels. (a) Car I and (b) car II. (c)(d) The reconstructed complex images of toy cars at 880mm from phaseshiftingholographic fringe pattern with size of 256×256 pixels. (c) Car I and (d) car II.
The total computation time measured by the CPU orGPU computing to differentiate the complex image of thereference object (car I) from that of the unknown input (carII). The size of complex image reconstructed by using aphaseshifting digital holography is varied from 256×256 to1152×1152 pixels.
256×256 pixels to 1152×1152 pixels. The computed crosscorrelation peak values between the reference (car I) holographic image and unknown input (car II) one were 0.0014,0.0010, 0.0016, 0.0011, 0.0008, 0.0004, 0.0003 and 0.0002,corresponding to holographic image sizes from 256×256 pixels to 1152×1152 pixels. The correlation peak values are normalized to the corresponding autocorrelation value of the reference car I holographic image. In the experiments, the size of the complex image reconstructed from the holographic fringe pattern is equal to that of the holographic fringe pattern. It is noted that the computed crosscorrelation peak values decrease as the complex image size increases.
Figure 5
shows the total computation time measured by the CPU or GPU computing to differentiate the complex image of the reference (car I) from that of the unknown input (car II) by changing the size of the complex image reconstructed by using phaseshifting digital holography.The complex image size was varied from 256×256 to 1152×1152 pixels. As shown in
Fig. 5
, the total computation
The singleexposure inline holographic fringe patternsof the toy cars with size of 1024×1024 pixels. (a) Car I and (b)car II.
time measured by the CPU computing to differentiate between car I and car II rapidly increased as the complex image size increased, while the equivalent by the GPU computing increased very slowly. It is noted that the size of the reconstructed complex image is the key factor to the computation time for both CPU and GPU computing.The computation time increases as the size of the complex image is increased. It is also noted that as the size of the complex image increases, the difference between the computation time between CPU and GPU computing increases rapidly. This demonstrates that parallel processing on a GPU has an enormous advantage over sequential processing on a CPU if the complex image size is large enough. As shown in
Fig. 5
, the increase in speed for the total computation time of 3D object recognition with the complex image size of 256×256 is approximately 8.5 while for the complex image size of 1152×1152, the increase is approximately 43.4.
Similarly, for singleexposure inline digital holography,we recorded the single interference patterns of each 3D car with a CCD camera with 2048×3072 pixels and pixel size of 12 ㎛×12 ㎛ using an inline interferometry setup as shown in
Fig. 2
. The singleexposure inline holographic fringe patterns of toy car I and car II were obtained by using Eq. (6), where we generated eight singleexposure inline holographic fringe patterns with different sizes ranging from 256×256 to 1152×1152 pixels for each car, in order to see their influence on the computational speed of 3D recognition for CPU or GPU computing.
Figure 6
shows
(a)(b) The reconstructed complex images of toy cars at 880mm from the singleexposure inline holographic fringe patternwith size of 1024×1024 pixels. (a) Car I and (b) car II. (c)(d) The reconstructed complex images of toy cars at 880mm from thesingleexposure inline holographic fringe pattern with size of 256×256 pixels. (c) Car I and (d) car II.
the singleexposure inline holographic fringe patterns with the size of 1024×1024 pixels for car I and car II.
For 3D imaging, the complex images of the car I and the car II were numerically reconstructed from the corresponding singleexposure inline holographic fringe pattern at a distance of 880mm by discrete Fresnel transform shown in Eq. (8)using CPU or GPU computing.
Figure 7
shows the reconstructed complex images of car I and car II from the corresponding singleexposure inline holographic fringe patterns.
For 3D recognition, the matched filter in Eq. (9) was applied to the reconstructed complex images of car I and car II using CPU or GPU computing, where the complex image size was varied from 256×256 to 1152×1152 pixels.The computed crosscorrelation peak values between the reference (car I) complex image and unknown input (car II)were less than 0.00005 in the range of the complex image size (256×256, 1152×1152). The correlation peak values are normalized to the corresponding autocorrelation value of the reference car I complex image.
Figure 8
shows the total computation time measured by the CPU or the GPU computing to differentiate the reconstructed complex image of the reference (car I) from that of the unknown input (car II) by changing the size of the singleexposure inline holographic image. The reconstructed complex image size was varied from 256×256 to 1152×1152 pixels. As shown in
Fig. 8
, the total computation time measured by the CPU computing to differentiate between car I and car II increases rapidly as the complex image size is increased,while the equivalent time by GPU computing increases slowly.It is noted that the size of the complex image reconstructed by using singleexposure inline digital holography is the key factor to the computation time for both CPU and GPU computing. The computation time increases as the size of the complex image is increased. It is also noted that as the size of the complex image increases, the difference of the computation times between CPU and GPU computing increases rapidly. It demonstrates that parallel processing on the GPU is much better than sequential processing on the CPU if the complex image size is large. As shown in
Fig. 8
, the speed advantage for the total computation time of 3D object recognition with the complex image size of 256×256 is approximately 9.0 while for the complex image size of
The total computation time measured by the CPU orGPU computing to differentiate the reconstructed compleximage of the reference (car I) from that of the unknown input(car II) using a singleexposure inline digital holography. Thereconstructed complex image size was varied from 256×256to 1152×1152 pixels.
1152×1152, the speed advantage is approximately 36.7. It is interesting to note that for both phaseshifting and singleexposure inline holographic fringe patterns, the computational time spent in differentiating the reconstructed complex image of the reference from that of the unknown input can be implemented in 21.5 ms~128.5 ms, which makes the 3D object recognition rate reach 7~46 frames/s. Therefore, it can be possible to obtain high object recognition rate faster than the normal realtime level (25 frames/s) by appropriately choosing the complex image size.
For the case of the 3D recognition of the object located at the same distance from the CCD camera, the matched filter in Eq. (10) can be directly applied to the phaseshifting or singleexposure inline holographic fringe patterns of two toy cars (car I and II) without complex image reconstruction.The computed crosscorrelation peak values between the car I with phaseshifting holographic fringe pattern and car II were less than 0.00007 with the holographic fringe pattern size varying in the range of (256×256, 1152×1152)
Results using phase shifting holographic fringepattern. The total computation time measured by the CPU orGPU computing to differentiate between the reference (car I)from that of the unknown input (car II) by changing the sizeof the phaseshifting holographic fringe pattern. Theholographic fringe pattern size was varied from 256×256 to1152×1152 pixels.
pixels. The computed crosscorrelation peak values between the car I and car II with singleexposure inline holographic fringe patterns were less than 0.00005 in the range of(256×256, 1152×1152). For comparison, the computed crosscorrelation peak values are normalized to the corresponding autocorrelation value of car I with phaseshifting holographic fringe pattern or singleexposure inline pattern.
Figure 9
show the total computation time measured by the CPU or the GPU computing to differentiate the phaseshifting holographic fringe pattern of the reference (car I)from that of the unknown input (car II) by changing the size of the phaseshifting holographic fringe pattern. As shown in
Fig. 9
, the total computation time measured by the CPU computing to differentiate between the car I and the car II rapidly increases as the holographic fringe pattern size is increased, while that by the GPU computing is increased slowly. It is noted that the size of the phaseshifting holographic fringe pattern is the key factor for time consumption for both CPU and GPU computing, which increases as the size of the holographic fringe pattern increases. It is also noted that as the size of the holographic fringe pattern increases, the difference between the computation time between CPU and GPU computing increases rapidly.This demonstrates that parallel processing on the GPU can increase the processing speed over sequential processing on the CPU if the phaseshifting holographic fringe pattern size is large.
Figure 10
show the total computation time measured by the CPU or GPU computing to differentiate between the reference (car I) from that of the unknown input (car II)by changing the size of the singleexposure line holographic fringe pattern. As shown in
Fig. 10
, the total computation
Results using singleexposure inline holographicfringe pattern. The total computation time measured by theCPU or GPU computing to differentiate between thereference (car I) from that of the unknown input (car II) bychanging the size of the singleexposure inline holographicfringe pattern. The holographic fringe pattern size was variedfrom 256×256 to 1152×1152 pixels.
time measured by the CPU computing to differentiate between the car I and the car II rapidly increases as the holographic fringe pattern size is increased, while that by the GPU computing increases slowly. It is noted that the size of the singleexposure inline holographic fringe pattern is the key factor to the computation time for both CPU and GPU computing which increase as the size of the holographic fringe pattern increases. It is also noted that as the size of the holographic fringe pattern increases, the difference between the computation time between CPU and GPU computing increases rapidly. It demonstrates that parallel processing on the GPU has an enormous speed advantage over sequential processing on the CPU if the singleexposure inline holographic fringe patterns size is large.
These experimental results in
Fig. 5
,
8
,
9
and
10
demonstrate that the proposed phaseshifting or singleexposure inline digital holography using GPU computing can provide significantly improved computational speed for realtime 3D object recognition as compared to CPU computation.
VI. CONCLUSION
We have investigated a realtime 3D object recognition system based on digital holography using a GPU chip. It has been illustrated that the computational speed of this 3D recognition system can be significantly faster than that of conventional CPU computing. Also, we have examined and compared the computation time of the 3D object recognition by use of digital holography with CPU and GPU computing by varying the size of the reconstructed complex image or the holographic fringe pattern of the 3D object. It has been shown as the size of the recorded holographic fringe pattern increases, the computation time of the 3D object recognition system with CPU computing increases rapidly, while that of the GPU computing increases slowly. These experimental results demonstrate that a 3D recognition system by parallel processing on a GPU has an enormous speed advantage over sequential processing on a CPU particularly for large holographic fringe pattern size. We have demonstrated that a digital holographic recognition system with GPU computing can provide significantly improved computational speed for realtime 3D object recognition as compared with CPU computing.
Acknowledgements
This study was supported by research fund from Chosun University, 2011.
Goodman J. W
,
Lawrence R. W
1967
Digital image formation from electronically detected holograms
Appl. Phys. Lett.
11
77 
79
Goodman J. W
Introduction to Fourier Optics
McGrawHill
New York USA
Yamaguchi I
,
Zhang T
1997
Phaseshifting digital holography
Opt. Lett.
22
1268 
1270
Kreis T
2005
Handbook of Holographic Interferometry
Wiley
New York USA
Martinez L
,
Javidi B
2008
Synthetic aperture singleexposure onaxis digital holography
Opt. Express
16
161 
169
Dubois F
,
Joannes L
,
Legros J.C
1999
Improved threedimensional imaging with digital holography microscope using a partial spatial coherent source
Appl. Opt.
38
7085 
7094
Abdelsalam D
,
Baek B
,
Cho Y
,
Kim D
2010
Surface form measurement using single shot offaxis Fizeau interferometry
J. Opt. Soc. Korea
14
409 
414
Ferraro P
,
De Nicola S
,
Coppola G
,
Finizio A
,
Alfieri D
,
Pierattini G
2004
Controlling image size as a function of distance and wavelength in Fresneltransform reconstruction of digital holograms
Opt. Lett.
29
854 
856
Cuche E
,
Marquet P
,
Depeursinge C
1999
Simultaneous amplitude and quantitative phase contrast microscopy by numerical reconstruction of Fresnel offaxis holograms
Appl.Opt.
38
6994 
7001
Zhang Y
,
Pedrini G
,
Osten W
,
Tiziani H. J
2004
Reconstruction of inline digital holograms from two intensity measurements
Opt. Lett.
29
1787 
1789
Nomura T
,
Javidi B
,
Murata S
,
Nitanai E
,
Numata T
2007
Polarization imaging of a threedimensional object by use of onaxis phaseshifting digital holography
Opt. Lett.
32
481 
483
Javidi B
,
Tajahuerce E
2000
Three dimensional object recognition by use of digital holography
Opt. Lett.
25
610 
612
Javidi B
,
Kim D
2005
Threedimensionalobject recognition by use of singleexposure onaxis digital holography
Opt.Lett.
30
236 
238
Lee H
,
Kim S
,
Kim D
2010
Two step onaxis digital holography using dualchannel MachZehnder interferometer and matched filter algorithm
J. Opt. Soc. Korea
14
363 
367
Javidi B
,
Moon I
,
Yeom S
,
Carapezza E
2005
Threedimensional imaging and recognition of microorganism using singleexposure online (SEOL) digital holography
Opt.Express
13
4492 
4506
Stern A
,
Javidi B
2007
Theoretical analysis of threedimensional imaging and recognition of microorganism technique using singleexposure online (SEOL) digital holography
J. Opt. Soc. Am.
A 24
163 
168
Shimobaba T
,
Sato Y
,
Miura J
,
Takenouchi M
,
Ito T
2008
Realtime digital holographic microscopy using the graphic processing unit
Opt. Express
16
11776 
11781
Moon I
,
Daneshpanah M
,
Javidi B
,
Stern A
2009
Automated threedimensional identification and tracking of micro/nanobiological organisms by computational holographic microscopy
Proc. IEEE
97
990 
1010
Shimobaba T
,
Sato Y
,
Miura J
,
Takenouchi M
,
Ito T
2008
Realtime digital holographic microscopy using the graphic processing unit
Opt. Express
16
11776 
11781