Object Oriented Framework for CUDA based Image Processing

Object Oriented Framework for
CUDA based Image Processing

組員名單
:

P76004588


徐華煊

P76004423

曾郁凱

P76014216

吳品頡





Keywords
:
Object
oriented
framework, CUDA,

design patterns,
image processing

Pritam

Prakash

Shete
,
Venkat

P. P.
K.,
Dinesh M.
Sarode
,
Mohini

Laghate
,
S. K.
Bose
& R. S.
Mundada
,

Bhabha

Atomic Research Centre, Mumbai, India

International
Conference on Communication, Information & Computing Technology (ICCICT
)
Oct.
19
-
20,2012,
Mumbai, India

1

1. Introduction

•
Compute Unified Device Architecture (CUDA)

–
CUDA is a novel and promising GPU programming frame work from NVIDIA.

–
The CUDA has been speedup many computationally intensive graphics as well
as nongraphic

•
Essential
for a seamless panoramic
mosaic

–
A
pyramidal image blending
algorithm

•
Our goal

–
T
o show
that use of design patterns facilitate extending existing functionality
by adding new classes, rather than modifying an existing classes or
functionality


2

1. Introduction

-

Panoramic Mosaic

3

1.
Introduction

-

Finding Key Point

4

2.
Analysis and Design

-

Gaussian
Blur
Operation

5

2
2
2
2
)
(
2
exp
2
1
)
,
,
(



y
x
y
x
G



)
,
(
*
))
,
,
(
)
,
,
(
(
)
1
(
y
x
I
k
y
x
G
k
y
x
G
n
n





)
,
,
(
)
,
,
(
)
1
(





n
n
k
y
x
L
k
y
x
L
)
,
(
*
)
,
,
(
)
,
,
(
y
x
I
y
x
G
y
x
L





doubles for

the next octave

)
,
,
(
)
1
(


n
k
y
x
D
L
(
σ
)

s
n
,
,
2
,
1


D(
σ
)

High frequency

k
= 2
(1/
s
)

s
: Image per octave

s
= 3 in this case

σ=
1.
6

σ=k
3
*1.
6

DoG filter

Gaussian filter

σ=k
4
*1.
6

σ=k
-
1
*1.
6

σ=k*
1.
6

σ=
k
2
*1.
6

σ=
1.
6

σ
=
2
*
1.
6

2. Analysis and Design

-

Laplacian

Pyramid

6

2. Analysis and Design

-

Remove
E
dges

Response

7

3. Implementation


-

Modules

•
Image Blending Library (IBL):
They developed framework for
CUDA based image processing.


•
This frame work offers
3 modules
for an image processing:

1)
CPU Module

2)
Simple
-
CUDA Module

3)
IO
-
CUDA Module






8

3. Implementation


-

Modules

1)
CPU Module

–
Using single thread


–
Implementation function:

1.
Gaussian blur

2.
Laplacian

pyramid

3.
REDUCE operation

4.
EXPAND operation





9

3. Implementation


-

Modules

2)
Simple
-
CUDA Module






CPU

GPU

Global
M
emory

CPU
Memory

Send

image

Process

Image

Return

image

10

3. Implementation


-

Modules

3)
IO
-
CUDA Module





GPU

Global
M
emory

Image already get

Shared
Memory

Process

Image

Send

image

Return

image

11

3. Implementation


-

Automatic Image Conversion

•
Visitor Design pattern

–
Image type
:
CPUBuffer

image and
CUDABuffer

image








Visitor

Concrete Visitor

Concrete Visitor

Element

Concrete Element

Concrete Element

Output:
CUDABuffer

Output:
CPUBuffer

12

3. Implementation


-

Image Source Integration

•
General Hierarchy Pattern




13

3. Implementation


-

Extensible Architecture

•
Construction of Gaussian &
Laplacian

Pyramid

14

3. Implementation


-

Extensible Architecture

•
Building the Gaussian
pyramid

–
Non
modifiable code along with placeholders for extending
it

•
Using
Factory
Method pattern

placeholders

–
C
reating
the image
buffer

–
Gaussian blur

–
REDUCE operation

•
Realized
by the respective
subclasses

–
Ex.
CPUBlendingFactory

–
Ex.
IOCUDABlendingFactory

15

3. Implementation


-

Extensible Architecture

•
Combine Pyramid Operation

16

3. Implementation


-

Extensible Architecture

•
Combine
P
yramid
O
peration

–
Validating
an input
pyramids

–
Allocating
memory for an output
pyramid

–
Combining
high pass and low pass
images

•
The
Template
Method pattern

is used to define the
skeleton

•
Combining
high pass images and low pass image
are
placeholder
or

extension
points

•
Subclasses
redefine

combining high pass and low pass
images without changing basic algorithm
structure

–
Ex.
CPUCombinePyramidOperation

–
Ex.
IOCUDACombinePyramidOperation


17

4. Result

•
Specification

–
Intel Core 2 Duo with E8400 3.00 GHz
processor

–
2GB
RAM

–
NVIDIA’s
Quadro

FX
4600

–
Input images covers
low resolution (128x128)
as well as a
high resolution (2048x2048)


18

4. Result

•
Panoramic
image
stitching

19

4. Result

•
CUDA based modules perform much better than the CPU
module

•
Simple
-
CUDA

–
Device global
memory

•
IO
-
CUDA module

–
Fast
on
-
chip shared memory


20

5. Conclusion

•
I
mplemented
an object oriented framework for a
GPU based image
processing

•
Using
software engineering principles and design
patterns

•
E
xtending
t
he
framework for computation using the
GPU
memory

21



THANK YOU

22