Gpu based image segmentation using

GPU-BASED IMAGE SEGMENTATION USING
LEVEL SET METHOD WITH SCALING
APPROACH
Zafer Güler1andAhmet Çınar2
1

Department of Software Engineering, Firat University, Elazig, Turkey
zaferguler@firat.edu.tr

2

Department of Computer Engineering, Firat University, Elazig, Turkey
acinar@firat.edu.tr

ABSTRACT
In recent years, with the development of graphics processors, graphics cards have been widely
used to perform general-purpose calculations. Especially with release of CUDA C
programming languages in 2007, most of the researchers have been used CUDA C
programming language for the processes which needs high performance computing.
In this paper, a scaling approach for image segmentation using level sets is carried out by the
GPU programming techniques. Approach to level sets is mainly based on the solution of partial
differential equations. The proposed method does not require the solution of partial differential
equation. Scaling approach, which uses basic geometric transformations, is used. Thus, the
required computational cost reduces. The use of the CUDA programming on the GPU has taken
advantage of classic programming as spending time and performance. Thereby results are
obtained faster. The use of the GPU has provided to enable real-time processing. The developed
application in this study is used to find tumor on MRI brain images.

KEYWORDS
CUDA, GPU Programming, Level Set Method, Image Segmentation, Scaling

1. INTRODUCTION
The first and most important step in image analysis is image segmentation. Segmentation is the
process of dividing an image that created it, and the computer vision [1] and medical imaging [2]
are among the most significant transactions. Another purpose of segmentation is to extract the
boundaries of the object or objects in the images. For this purpose, level set method is used to
extract object boundaries in images with highly successful results. The level set method is
released by Osher and Sethian [3]. After it is released for the first time, it has become a favorite
technique for finding, tracking or monitoring of moving surfaces. The overall objective of the
level set method is to monitor moving surfaces and curves. Here, the moving surfaces can be
traced by the speed term and the speed term depends on the current geometry. By means of this
method, topological changes automatically manageable and an object can be divided into more
than one object, or vice versa.

Sundarapandian et al. (Eds) : ICAITA, SAI, SEAS, CDKP, CMCA-2013
pp. 81–92, 2013. © CS & IT-CSCP 2013

DOI : 10.5121/csit.2013.3808

Computer Science & Information Technology (CS & IT)

82

Level set method quite suitable for the method such as mud and water physically-based
simulations [4], but the method require solving the partial differential equations (PDEs), so the
solution requires a high computational cost. Many methods have been developed to deal with this
problem. Two major approaches are narrow band [5] and sparse field technique [6]. With these
methods, PDE improvements calculated only around the zero level set and the speed
improvement has obtained. But enough acceleration could not achieve. Later studies focused on
running parallel to the level set method. Thus the parallel algorithms have been developed [7].
Multiprocessor structures emerged in the last ten years, and today began to be used widely. These
structures are used in both industrial and academic research, and the difficulties and problems can
be solved with these structures. Nowadays, many problems can be implemented easily with this
structure. Especially with the rapid advances in video card, running parallel to the level set
method has become an important factor to the speed increase.
Graphics processor unit (GPU) to be used for general purpose applications is not an approach that
emerged in recent years, but in 2007, with the development of the CUDA architecture, it is
expanded rapidly. The CUDA architecture provide to execute general purpose application without
knowledge of the graphics processor. GPU-based applications are used not only in the scientific
field, but also in other fields that require high performance such as, image and video processing,
fluid dynamics simulation [8].
Level set method began to run on the GPU in the early 2000s. The first GPU implementation of
the level set method is implemented by Rumpf and Strzodka[18] in 2001. Researchers continued
to study in this area until 2007[19, 11]. With the release of CUDA C programming language in
2007, high performance solutions have been obtained [12-15]. One of the important studies on the
level set method was presented by Robert et al. [16] and Jalba et al [17]. Robert’s method
represents approach to implement narrow-band approaches on the GPU and Jalba’s method
represents approach to implement sparse approaches on the GPU.
In this paper, a novel GPU-based level set method is presented. The proposed method does not
include PDE solution, so our algorithm has significantly increased the speed of calculation. We
use basic geometric transformations for curve evolution. This paper is organized as flows. In
Section 2 provides information on the new generation of GPU and memory management. In
section 3 we explain the major details of developed GPU-based algorithm. Finally, results
andtime measurements are presented.

2. CUDA ARCHITECTURE
Graphics processors are rapidly developed and become available for general purpose applications.
Thus it has become to use in many areas to improve the speed of application. CUDA architecture
developed by NVIDIA company and this architecture allows high increase in computing
performance. In fact, this performance difference is due to the fact that, the GPU architecture
designed for parallel operations. A simple CPU (central processor unit) and GPU architecture are
illustrated in figure 1. As shown in the illustration GPU has a large number of arithmetic logic
unit (ALU), but its cache memory is low [18].

83


Figure 1.CPU and GPU architecture [18].

GPUs execute a large number of threads on a set of data at the same time. Therefore it is only
appropriate for parallel data. Thus the successful result can be obtained. For example if a program
contains many flow control, its calculation speed may be reduce rather than increase [7].

2.1. CUDA C Programming Language
A CUDA C programming language introduces a small number of extensions to C language and
allows us to define C function called kernel. With kernel we can define and execute N function in
parallel. When kernel calls, unique threadId is available. Thus, with the Id number we can
determine which thread is currently running within the kernel. Threads are organized in blocks. In
the same as thread, blocks have an identification number. So that which block and which thread
actually running, can be easily determined in large data sets. But the number of thread in the
block is limited. Current GPUs are supported up to 1024 thread. Threads and blocks can be one,
two or three dimensional. Blocks are organized in grids. A general structure is illustrated in figure
2 [18].

Figure 2.Grid of the blocks [18].


84

2.2. Memory Management
CUDA architecture supports many memory types. Threads, blocks and type of memory are
illustrated in figure 3. Each thread access registers, local memory, shared memory, constant
memory and global memory. The CPU part of the application can be access global memory and
constant memory. The used memory types are listed below.
Registers: Registers are defined in the thread. These variables cannot be accessed from outside
the thread. Generally, it is used to store local variable in the function. It is not require any extra
programming extension [8, 9, 10]
Local Memory: Same as registers it is valid only in the thread. It is defined by __local__
keyword. t is slower than register [8].
Shared Memory: All of the thread in the blocks can access to shared memory. The values in the
blocks can be accessed by any thread. In general, it is fast as registers. t can be defined by
__shared__ keyword [8, 9].
Global memory: All application (CPU and GPU part) can access global memory. It is defined by
__device__ keyword. Data transfer speed is very slow [18].
Constant Memory: Constant memory can only be read by the GPU. Each running thread can read
from memory at the same time. Thus, we obtain very fast data transfer [8, 10].

Figure 3.GPU hardware architecture and memory types [20].

85


3. DEVELOPED METHODS
In this paper, a novel level set method based on scaling approaches for extracting object on the
images is proposed. In this section, developed GPU algorithm will be discussed. The application
is implemented by using the Visual Studio 2012 environment with CUDA C language. The
developed method consists of the following five stages. These stages and their working devices is
presented in Figure-4.

Figure 4.General algorithm of the developed method.

3.1. Preparation Stage
In this stage, user can select any picture by means of the developed user interface. After getting
the image from a user, it sends to the graphics processor. OpenGL library was used to display the
picture from the user. This process will be performed by the CPU.

3.2. Pre-processing Stage
In this stage some preprocessing techniques are performed on the images. The first technique is
noise reduction. We use Gaussian smoothing for this purpose [21]. The Gaussian smoothing
operator is a 2-D convolution operator that is used to remove detail and noise. We use 5x5
Gaussian filter. The second technique is to convert color to grayscale. The developed method
applies on gray level images so gray level transformation is made in this section for color images.
The final technique is Sobel operator. This operator is used in image processing, particularly edge


86

detection algorithm. This technique is used to help “EDGE” algorithm which is mentioned in
section 3.4. These three techniques executes on the GPU.

3.3. Determination of Contour and Center
Sometimes the initial contour is very important to find object boundaries quickly. The developed
application has tested on MRI brain tumors images. Here, we have expected to identify tumor
area sketchy by expert. Determined initial contour curve is drawn as a circle. Then this curve is
sent to the graphic processor to provide curve evolution.
The center of the object is the main element of the curve evolution algorithm. But we have no
idea about the center of the object. Fortunately, it can be adjusted by the system automatically or
by an expert. In this paper, we prefer to determine the center by an expert. Thus, we provide a
more accurate center. Also, we allow selecting more than one center, so that the concave shapes is
successfully determined. An example is illustrated in figure 5. U-shaped object having a concave
region at the top of the figure. Figure 1a shows initial condition, figure 1b shows the result of
boundaries with one center, and figure 1c shows the result of boundaries with three centers. As
shown in the figure, using more than one center is more accurate.

(a) Initial Condition

(b) Result with one center

(c) Result with three centers

Figure 5.Comparison of the use of a single center and multi-center on U-shaped object.

3.4. Contour Evolution
In level set method, curve evolution is based on the solution of partial differential equation. This
causes the algorithm to run slowly. The developed algorithm for this paper use different approach
for curve evolution.
The developed curve evolution method uses two-dimensional geometric transformations such as
scaling and it uses the logic of binary search. Binary search technique is applied on series of
numbers and with this techniques search time significantly reduces. In developed algorithm, the
evolution of next point on the curve is determined for each point. Normally, binary search is
performed on numbers but in our method, we use pixels for binary search operation. Rather than a
set of number of binary search, we define a set of pixels and half of the cluster is eliminated at
each cycle approximately. This process continues until the edge is found or until there are no
elements in the cluster. Elements of the cluster are pixels between a point on the contour and
shape center.

87


Figure 6.The movement of a pixel on the curve. (a) Initial state. (b) After first step. (c) After secondstep. (d)
After third step. (e) After fourth step – Edge found.

Algorithm is illustrated for a pixel in Figure 6. Thus, when we move contour pixel by pixel, the
running time of the edge detection process will be O(n) , but with this algorithm the running time
of the edge detection process will be O(logn) (n is the number of pixels between initial point and
center point). This process is repeated for all points on the curve, by this way the development of
the initial curve is provided.
In algorithm, we use scaling in order to ensure the movement of pixels. At the beginning of the
evolution, there are two variables defined for scaling which are called “sr1” and “sr2” and
respectively they are set 1 and 2. “sr1” variable refers to scaling factor which is applied to the
previous cycle and “sr2” variable refers to current scaling factor. Variables are calculated at the
end of each cycle by the following formula according to the movement of each pixel.
‫2ݎݏ = 1ݎݏ‬

|‫|2ݎݏ − 1ݎݏ‬
2
‫ = 2ݎݏ‬൞
|‫|2ݎݏ − 1ݎݏ‬
‫− 2ݎݏ‬
2
‫+ 2ݎݏ‬

(1)
‫ݐ ݂݋ ݁݀݅ݏ݊݅ ݈݁ݔ݅݌ ݂ܫ‬ℎ݁ ‫ݏ‬ℎܽ‫݁݌‬

‫ݐ ݂݋ ݁݀݅ݏݐݑ݋ ݈݁ݔ݅݌ ݂ܫ‬ℎ݁ ‫ݏ‬ℎܽ‫݁݌‬

(2)

Developed GPU algorithm is given in Figure 7. Algorithm uses two functions. The first function
is “EDGE” function. This function provides information about whether or not a pixel closes to the
edge. The other function is “INSIDE” function. This function is used to determine a pixel is
inside or outside the shape.


88

Figure 7.GPU algorithm of the algorithm.
Figu

“EDGE” function uses the calculated edge information which is calculated in preprocessing
section and return in the form of true or false. In the same way “INSIDE” function returns result
as true or false. “INSIDE” function uses centers values, initial contour and the average density of
nction
contour area to decide. Every pixels on the contour execute on a CUDA thread, so application
accelerated. Curve evolution algorithm is applied an MRI liver image and illustrated in figu 8.
figure
The image size is 512x512 and as shown in the illustration contour evolution is very fast.

(a) Initial State

(b) Cycle-1

(c) Cycle
Cycle-2

(d) Cycle-3

(e) Cycle-5

(f) Cycle
Cycle-7

Figure 8.An example of curve evolution algorithm.
8.

3.5. Presentation
At this stage, the entire calculation has been completed. Final contour and preprocessed image
copy from GPU to CPU memory. Then, contour points and image are combined and presented to
the user.

89


4. RESULTS
In this section we present experimental results obtained by the proposed methods. All
experiments were performed on a machine equipped with an Intel Core i7-3770 CPU at 8GB
RAM and a GeForge GTX 660 Ti GPU. GPU has 7 streaming multiprocessor (SM) and each SM
has 192 CUDA processor. This means 1344 computing core per chip. The number of register per
multiprocessor is 65536, the total amount of constant memory is 64KB. The amount of shared
memory per multiprocessor is 48KB, organized into 32 banks. The 2GB amount of global
memory is reached through a GGDR5 interface. The architecture supports the double precision
floating point arithmetic.
In this study, we use MRI brain tumor images. The resolution of the image is512x512. The results
are illustrated in figure 9. As shown in the figure the tumor has found with high success rate.

Figure 9.The results of the proposed method.


90

The developed application has been tested with different settings on different images with
different resolution. Firstly, the preprocessing operation is tested with different settings. We test
our algorithm with three different parameters. These are block number, thread number and image
resolution. The results are shown in Table 1. Table 1 also shows that execution time generally
depends on the image resolution but block number and thread number are also very important for
performance optimization. As shown in second and last row of the table 1, thread number is
selected very low, so the execution speed increases.
Table 1. Execution time of the preprocessing stage.

Image Resolution Block Number Thread Number Execution Time (ms)
256x256

16x16

16x16 (256)

0,36

256x256

32x32

8x8 (64)

0,70

256x256

8x8

32x32 (1024)

0,36

256x256

16x8

16x32 (512)

0,36

512x512

16x16

32x32 (1024)

1,34

512x512

32x32

16x16 (256)

1,33

512x512

16x32

32x16 (512)

1,32

512x512

32x64

16x8 (128)

1,33

1024x1024

64x64

16x16 (256)

5,14

1024x1024

64x32

16x32 (512)

5,19

1024x1024

32x32

32x32 (1024)

5,19

1024x1024

128x64

8x16 (128)

5,36

Secondly, curve evolution algorithm is tested with different settings. We test our algorithm with
four different parameters. These are block number, thread number, image resolution and number
of contour points. The results are shown in Table 2. Considering that, curve evolution algorithm
is a circular structure, so we only give one cycle’s execution time. In Table 2, execution time is
given in milliseconds and indicates that, the slow-running cycle. As shown in table 2, the
execution time is not change with the image resolution, because we use image only for
information. In other word, we do not use image information primarily, so threads will access
image data with global memory when it is necessary.. We use contour points as primary data, but
as shown in the figure 2, it is generally no effect for execution time. There are two reasons for this
result. Firstly, the complexity of the algorithm is O(1). Each thread executes their code and this
code does not include a circular part. Second and main reason, the video card capacity is not used
in this process. As shown in table 2, threads and blocks number are very limited. Here, the
important factor for execution time is to select a sufficient number of threads. For example, the
second, fourth and sixth row of the table 2, we select 4x4 thread number. But there is no effect for
performance. On the other hand, when we increase the number of contour points, 4x4 thread
number is not enough and this leads to increase execution time. For example, in seventh row in
table 2, we use 4096 contour points and 4x4 thread number is not enough.
A key decision for performance optimization in CUDA programming is not only the choice of the
size of the block and thread number, but also the choice of the memory type. With an only small
change to the type of memory used, can be achieved great acceleration. We use shared and
constant memory in this study. The main aim is to maximize memory bandwidth. For this,in

91


preprocessing stage, we use shared memory for image convolution. Also the constant memory is
used in all GPU-based function.
Table 2. Execution time of the curve evolution algorithm. (For 1 cycle)

Image
Resolution
256x256

Block
Number
4x4

Thread
Number
8x8

Contour Point
Number
1024

Execution Time
(ms)
0,058

256x256

8x8

4x4

1024

0,055

256x256

4x2

4x8

256

0,057

256x256

4x4

4x4

256

0,054

512x512

4x4

8x8

1024

0.058

512x512

8x8

4x4

1024

0.056

512x512

16x16

4x4

4096

0.090

512x512

8x8

8x8

4096

0.059

1024x1024

4x8

16x8

4096

0.060

1024x1024

4x4

16x16

4096

0.060

5. CONCLUSIONS
In this paper, a novel GPU-based level set method has presented with scaling approach.We
followed the method of heterogeneous programming technique for this study. Some of the steps
were made on the CPU side and some of the steps were made on the GPU side. Minor procedures
and user inputs performed on the CPU side. Operations that require high computing such as curve
evolution and preprocessing performed on the GPU side. Thus, we achieved very fast execution
time.
In developed method, the curve evolution algorithm used basic geometric transformation like
scaling, so that the algorithm has been run much faster. As a result, all operations were completed
in about 2 milliseconds with all memory operations. Thus, it is possible to show that with the
developed method, we can perform real-time operation.

REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]

S. Osher, and N. Paragios, (2003)“Geometric Level Set Methods in Imaging, Vision and Graphics”,
Springer-Verlag New York, Secaucus, NJ, USA.
D. W.Shattuck, G. Prasad, M. Mirza, K. L. Narr, and A. W. Toga, (2009)“Online resource for
validation of brain segmentation methods”, NeuroImaging, 45, 431-439.
S. Osher, and J. Sethian, (1998)“Fronts propagating with curvature-dependent speed: Algorithims
based on Hamilton-Jakobi formulation”, J. Comput. Phys., 79-1, 12-49.
R. Fetkiw, (2002)“Simulating Natural Phenomena for Computer Graphics, Geometric Level Set in
Imaging”, Vision and Graphics, 461-479.
D. Adalsteinsson, and J. Sethian, (1995)“A Fast Level Set Method for Propagating Interfaces”,
Journal of Computational Physics, 118-2, 269-277.
R. Whitaker, (1998)“A Level-Set Approach to 3D Reconstruction from Range Data”, International
Journal of Computer Vision, 29-3, 203-232.

[7]

[8]
[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]
[18]
[19]
[20]
[21]

92

S. P. Awate, and R. T. Whitaker, (2004)“An Interactive Parallel Multiprocessor Levet-Set Solver with
Dynamic Load Balancing”, Scientific Computing and Imaging Institute Technical Report, UUCS-05002, University of Utah, Salt Lake City, UT, USA.
J. Sanders, and E. Kandrot, (2011)“CUDA By Example: An Introduction to General-Purpose GPU
Programming”, Addison-Wesley, NVIDIA Cooporation.
M. Rumpf, and R. Strzodka, (2001) “Level Set Segmentation in Graphics Hardware”, IEEE
International Conference on Image Processing (ICIP’01), Thessaloniki, Greece, October 7-10, pp.
1103-1106.
A. Lefohn, and R. Whitaker,(2002)“A GPU-Based Three-Dimensional Level Sey Solver with
Curvature Flow”, Scientific Computing and Imaging Institute Technical Report, UUCS-02-017,
University of Utah, Salt Lake City, UT, USA.
A. Lefohn, J.M. Kniss, C. D. Hansen, and R. T. Whitaker, (2004)“A Streaming Narrow-Band
Algorithm: Interactive Computation and Visualization of Level Sets”, IEEETransactions on
Visualization and Computer Graphics, 10, 422-433.
O. Klar, (2007)“Interactive GPU-based Segmentation of Large Medical Volume Data with LevelSets”, 11th Central European Seminar on Computer Graphics (CESCG’07), Budmerice Castle,
Slovakia, April 23 - 25.
H. Mostofi, and K. Colege, (2009)“Fast level Set Segmentation of Biomedical Images using Graphics
Processing Units”, Final Year Project, University of Oxford, Department of Engineering Science
Oxford.
A. Hagan, and Y. Zhao, (2009)“Parallel 3D Image Segmentation of Large Data Sets on a GPU
Cluster”, 5th International Symposium on Visual Computing, Las Vegas, Nevada, USA, November
30 - December 2, pp. 960-969.
G. J. Tornai, and G. Cserey, (2010)“2D and 3D level-set Algorithm on GPU”, 12th International
Workshop on Cellular Nanoscale Network and Their Applications (CNNA), Berkeley, CA, USA,
February 3-5, pp. 1-5.
M. Roberts, J. Packer, M. C. Sousa, and J. R. Mitchell, (2010)“A Work-Efficient GPU Algorithm for
Level Set Segmentation”, High Performance Graphics (HGP ‘10), Saarbrucken, Germany, June 2527, pp. 123-132.
A. C. Jalba, W. J. Van Der Laan, and J. B. T. M. Roerdink, (2013).“Fast Sparse Level Sets on
Graphics Hardware”, Visualization and Computer Graphics, 19-1, 30-44.
NVIDIA., (2012) “CUDA C Programming Guide v5.0”.
http://docs.nvidia.com/cuda/pdf/CUDA_C_Programming_Guide.pdf, 15 Mayıs 2013.
S. Cook, (2012)“Cuda Programming: A Developer’s Guide to Paralel Computing with GPUs”,
Elsevier, Morgan Kaufmann.
J. Ooster,“Cuda Memory Model”, http://3dgep.com/?p=2012, 14 Mayıs 2013.
“Gaussian Smoothing”, http://homepages.inf.ed.ac.uk/rbf/HIPR2/gsmooth.htm, 14 Mayıs 2013.

Authors
ZaferGüler is a research assistant in the Department of Software Engineering at the Firat
University of Elazig where he has been a faculty member since 2010. He completed his
master at Firat University Computer Engineering department. His research interests are
GPU programming, level set methods.image segmentation.
AhmetÇınar was born in Elazig (1972). He received the PhD degree in ElectricElectronics Engineering in 2003 from Firat University. He has graduated in 1993 BSc.
He has been working on Firat Univ. Department of Computer Engineering, (Assistant
Professor). His research is interested in development and improvement of mesh
generation methods, and applications of virtual reality, augmented reality, artificial
intelligence and game programming.

Gpu based image segmentation using

Recommended

More Related Content

What's hot (17)

Viewers also liked (20)

Similar to Gpu based image segmentation using (20)

Recently uploaded (20)

Gpu based image segmentation using