Dynamic Texture Coding using Modified Haar Wavelet with CUDA

Premanand P Ghadekar et al. Int. Journal of Engineering Research and Applications www.ijera.com
ISSN : 2248-9622, Vol. 5, Issue 3, ( Part -1) March 2015, pp.56-60
www.ijera.com 56 | P a g e
Dynamic Texture Coding using Modified Haar Wavelet with
CUDA
Premanand P Ghadekar*, Nilesh A Parmar**, Nilkanth B Chopade***
*(Department of Applied Electronics, SGBA University, Amravati-02)
** (Department of Electronics Engineering, VIT college, Pune-37)
***(Department of Electronics Engineering, PCCOE college, Nigadi, Pune-44)
ABSTRACT
Texture is an image having repetition of patterns. There are two types, static and dynamic texture. Static texture
is an image having repetitions of patterns in the spatial domain. Dynamic texture is number of frames having
repetitions in spatial and temporal domain. This paper introduces a novel method for dynamic texture coding to
achieve higher compression ratio of dynamic texture using 2D-modified Haar wavelet transform. The dynamic
texture video contains high redundant parts in spatial and temporal domain. Redundant parts can be removed to
achieve high compression ratios with better visual quality. The modified Haar wavelet is used to exploit spatial
and temporal correlations amongst the pixels. The YCbCr color model is used to exploit chromatic components
as HVS is less sensitive to chrominance. To decrease the time complexity of algorithm parallel programming is
done using CUDA (Compute Unified Device Architecture). GPU contains the number of cores as compared to
CPU, which is utilized to reduce the time complexity of algorithms.
Keywords – Compression, CUDA, Dynamic texture, GPU, Haar wavelet, PSNR, SSIM
I. INTRODUCTION
Video compression is a fast growing field with
many video compression algorithms are available.
The video contains highly correlated neighboring
pixels which are redundant. This redundancy can be
removed to achieve higher compression ratio [1].
The dynamic texture contains more redundant
part as compared to video. This redundant part can
be reduced. The dynamic texture gives high
compression ratio as compared to video. The visual
quality is also good as compared to video.
In the proposed algorithm, the 2-D modified
Haar wavelet is used. It is a simple and efficient
averaging differencing method applied in rows and
columns to achieve high compression ratios with
good visual quality.
GPU [2] is also called visual processing unit
(VPU). Originally GPUs were manufactured for
appropriate rendering of gaming graphics, but now
they are applied to scientific calculations also. GPU
has an extremely parallel array of integer and
floating-point processors with dedicated high-speed
memory. GPU has a parallel architecture that
executes many concurrent threads slowly instead of
executing a single thread very quickly. As shown in
fig.1, GPU is a co-processor to the CPU. Data and
functions should be off-loaded to the GPU from
CPU. Kernel is a function (code) written for
execution on the GPU, and it is executed as many
different threads in parallel. Core is a single
independent computational unit within a CPU or
GPU chip. CPU core performs general operations
whereas GPU core performs specialized operations.
CPU has four cores, and GPU has 240 to 2048 cores.
A GPU accelerates an application that satisfies the
following criteria: 1) The computations divided into
a large number of independent units of work. 2) The
time consumed in computation significantly
surpasses the time consumed in transferring data to
and from GPU memory. Programming is done in
CUDA [3] to reduce the time complexity of the
proposed algorithm. CUDA allows parallel
programming in CPU and GPU.
II. LITERATURE SURVEY
JPEG uses the Discrete Cosine transform which
gives a high compression ratio but have
disadvantages like blocking artefacts, graveness and
blurred video [4]. Traditional compression
algorithms have high time complexity. Traditional
algorithms only exploit spatial and temporal
correlation among pixels but chromatic correlation
among pixels are not used.
Scalable Accordion-DCT based video coding
method [5] converts 3D video into 2D structure. It
allows extracting the temporal redundancy of video
using up/down sampling method (SVC) which is
based on a combination of the forward and backward
DCT. Such technique prevents the motion
compensation step and decreases the number of
computations required. This method transforms
multiple frames into one large frame, and scalable
Accordion-DCT is applied on it. In Accordion-DWT
[6], 2D Daubechies wavelet transform applied to
RESEARCH ARTICLE OPEN ACCESS

large frame instead of up/down sampling method. It
overcomes the limitations of Accordion-DCT by
getting high compression ratio without much
affecting the number of computations.
The YCbCr color model is used to exploit
chromatic correlation among pixels; Cb and Cr
components down-sampled by half without loss of
visual quality as a change in chromaticity are not
visible to human eyes. 2-D Haar wavelet is simple
and efficient transform to achieve high compression
ratios without loss of visual quality. 2-D Haar
wavelet is computationally less complex as
compared to other transform as only averaging
differencing is used, and no temporary variable is
used during transform [7]. Traditional algorithms
have a high time complexity which reduced by
heterogeneous programming (CPU+GPU) in CUDA
[8].
III. Haar wavelet
Wavelet indicates “let the wave go” because
wavelet considers finite range of signal instead of its
infinite length. Fourier transform splits the signal
into a sequence of sine waves of different
frequencies, whereas the wavelet transform splits the
signal into its "wavelets", scaled and shifted versions
of the mother wavelet. The wavelet is irregular in
shape and compactly supported. These properties
make wavelets an ideal tool for analysing signals of
a non-stationary nature. Their irregular shape
enables them to analyse signals with discontinuity or
sharp changes, while their compactly supported
nature provides temporal localisation of a signal's
features. The wavelet transform provides localisation
of a signal in both time and frequency domain which
is very important to evaluate signals of the non-
stationary nature. The location of the wavelet allows
explicitly to represent the location of events in time.
The shape of the wavelet allows representing
different detail or resolution.
Alfréd Haar proposed the Haar wavelet in 1910.
It is irregular and not continuous. The Haar basis is
obtained with a multi-resolution of piecewise
constant functions. The scaling function is Φ=1[0, 1].
The filter h[n] =
has value 2-1/2
at
n=0 and n=1. Hence,
(1)
So,
The Haar wavelet is the only real compactly
supported wavelet that is symmetric or anti-
symmetric and which generates an orthogonal basis
of Hilbert space L2(R) [9]. Haar wavelet is simple
averaging and differentiating method which is
applied on either rows or column. The 2D-modified
Haar wavelet is same as Haar wavelet difference is
that it is applied on rows and then on the column. As
only averaging and differentiating is used the image
quality does not decrease [10].
The 2D-modified Haar wavelet [11] has
properties like it is real and orthogonal. The basis
vectors of the Haar transform are sequentially
ordered. The original signal is split into low and
high-frequency part. It has linear phase, compact
support, the and perfect representation properties.
IV. Encoder
Step 1: Take input as uncompressed video (.avi).
Step 2: Frames are extracted from Video. Each
Frame is converted from RGB to YCbCr.
Step 3: The Cb and Cr components are down-
sampled by factor 2. Down-sample by factor 2
means only considering alternate row and column of
the image matrix. Down sampling Cb and Cr
components does not affect the video quality
because human eyes are less sensitive to
Chrominance and its effect is negligible to the
human eye. Suppose Cb and Cr components have
512*512 pixel matrix then after down sampling it
becomes 256*256.
Step 4: Each Image component is divided into Bock
size of 8 by 8.Instead of applying DWT on the
whole image we use it on 8*8 blocks because
neighboring pixels are highly correlated.
Step 5: Algorithm to find out block redundancy is as
follows:
1. Apply DWT to 8*8 blocks.
2. Calculate summation of pixels in LL band and
store in variable say sum.
3. Apply DWT to other blocks, and store
summation of LL band pixels invariable say
sum1
4. If sum = sum1 +/- error, error=1,2,3,4,5 then
block is redundant
5. Else not redundant
If a block is redundant store only references of
that block i.e. frame no and block no. else perform
following steps.
Step 6: Algorithm for 2D-modified Haar wavelet
1. Find the mean of each pair of samples.
2. Find the difference between each average.
3. Replace the first half of the array by average
values.
4. Replace the second half of the array by
difference values.
5. Repeat the procedure on the first half of the
array.
Step7: Quantization is used to achieve lossy
compression, the algorithm is as follows:
1. Choose Delta value between 1 to 16
2. All pixels having values between minus Delta to
plus Delta are randomized to zero.

As the value of Delta increases compression
ratio increases and video quality decrease. Delta = 4
gives optimal results.
Step 8: Instead of applying RLC on rows or column,
applying on zigzag scanned matrix provides more
zero count. The advantage of zigzag scan is that we
access the neighboring pixels which are highly
correlated.
Step 9: It is applied to Zigzag scanned matrix. RLC
gives output in the form of pairs (Pixel value,
count).Here count means the no of times pixel value
repeated.
Step 10: The run length pairs stored in compressed
.txt File.
V. Decoder
Step 1: If the reference found in the compressed file,
then copy Run length values of that block.
Step 2: (Pixel Value, Count) converted into pixel
value and pixel value is copied count no of times.
Step 3: The block traversed in reverse zig zag style.
Step 4: Reverse algorithm of DWT is applied.
Step 5: Up sampling the Cb and Cr components by
factor 2.Take first row as output, take an average of
first and second row as output and take a second row
as output. Perform this for all rows and column.
Suppose Cb and Cr components have 512*512 pixel
matrix then after up sampling it becomes
1024*1024.
Step 6: YCbCr frames are converted to RGB frames.
Step 7: All frames are collected to get reconstructed
video.
VI. Parallel Programming
To achieve better performance implementation
is done in CUDA architecture. The use of GPUs for
computations and calculations provides a platform to
write applications that can run on a large number of
cores. CUDA is a NVIDIA’s employment of parallel
architecture [12] on their GPUs and offers APIs for
programmers to develop parallel applications using
the C programming language with some constraints
and CUDA extensions. CUDA delivers a software
abstraction for the hardware called grid. Grid is a
group of blocks which are a group of threads that
can share memory as depicted in fig.2. These threads
are then assigned to multiple cores (processors).
Eight scalar processors constitute a multiprocessor,
and different models of GPUs contain various counts
of multiprocessor.
Fig.1: CPU and GPU interfacing
Fig.2: Different memories of GPU
VII. Results
The proposed algorithm is implemented using
CUDA on different videos and dynamic textures.
CPU used for processing is 64-bit Intel Core i3 with
2.10GHz speed and 3 GB RAM. To parallelize code
GPU version Geforce GTX 525M with 96
processing cores and 1 GB dedicated graphics
memory is used. Dynamic textures with different
frame size are taken from Renaud Péteri et al.’s
dynamic texture database [13]. Parameters like
PSNR and SSIM are considered to compare visual
quality of original and reconstructed videos.
Compression Ratio (CR) represents the efficiency of
proposed algorithm with considerable visual quality.
Table-1 and table-2 display CR, PSNR and SSIM for
different videos without and with YCbCr
respectively. Fig.3 and fig.4 exhibit graphs of
original and compressed size for different videos
without and with YCbCr respectively. Table-3 and
table-4 represent CR, PSNR and SSIM for standard
dynamic textures without and with YCbCr
respectively. Fig.5 and fig.6 exhibit graphs of
original and compressed size for different textures
without and with YCbCr respectively.
Table 1: Compression ratio, PSNR and SSIM for
different RGB videos
Video Size
(MB)
Compressed
Size (MB)
Compression
Ratio
PSNR
(dB)
SSIM
United
(720*420)
838 115.10 86% 35.34 0.97
Bbb
(624*272)
773 157.75 79.59% 20.92 0.86
Hf
(608*240)
772 78.85 89.78% 21.48 0.86
Viptrain
(360*240)
154 49.05 68% 34.42 0.93
Sample
(512*512)
320 74.24 76.8% 28.12 0.85

Fig. 3: Original and compressed size of different
videos
different RGB videos with YCbCr
Video Size
(MB)
Compressed
Size (MB)
Compression
Ratio
PSNR
(dB)
SSI
M
United
(720*420)
838 74.88 91.06% 33.34 0.97
Bbb
(624*272)
773 104.03 86.54% 19.92 0.86
Hf
(608*240)
772 78.53 89.82% 20.38 0.86
Viptrain
(360*240)
154 32.35 78.99% 32.12 0.88
Sample
(512*512)
320 48.92 84.71% 28.12 0.85
videos with YCbCr
different dynamic textures
Video Size
(MB)
Compressed
Size (MB)
Compression
Ratio
PSNR
(dB)
SSIM
Steps
(355*288)
29.6 3.56 87.97% 42.77 0.97
Clouds
(320*288)
55.1 5.48 90.05% 36.99 0.971
Camel
(320*288)
55.1 5.15 90.65% 35.27 0.96
Bgrass
(360*240)
72.8 8.78 87.93% 37.67 0.965
Bird
(320*240)
23.2 4.96 78.62% 39.75 0.988
dynamic textures
different dynamic textures with YCbCr
Video Size
(MB)
Compressed
Size (MB)
Compressio
n Ratio
PSNR
(dB)
SSIM
Steps
(355*288)
29.6 2.34 92.04% 41.57 0.96
Clouds
(320*288)
55.1 3.62 93.43% 35.99 0.97
Camel
(320*288)
55.1 3.40 93.82% 34.07 0.95
Bgrass
(360*240)
72.8 5.79 92.04% 37.03 0.95
Bird
(320*240)
23.2 3.27 85.90% 38.05 0.97
dynamic textures with YCbCr

Table-5 shows comparison of time taken by CPU
only and CPU with GPU for different dynamic
textures. Fig. 7 demonstrates the same using graph.
Table 5: CPU and CPU+GPU time for different
dynamic textures
Dynamic
Texture
CPU
time(Sec)
CPU+GPU
time (Sec)
%ET
reduced
Bgrass 170.37 151.14 11.3
Chair 184.38 159.20 13.7
DNA 159.22 143.08 10.1
Grass_water 39.96 35.36 11.5
Duck_bush 133.58 118.93 10.9
Fig. 7: Comparison of CPU and CPU+GPU time
VIII. CONCLUSION
The 2D-modified Haar wavelet gives high
compression ratio with good PSNR ratio. The 2D-
modified Haar wavelet is simple and efficient
transform to implement which only have averaging
and differing, and no temporary array required for
transform. The proposed algorithm increases
compression ratios by using different quantization
levels.
The texture video contains more no of
redundant blocks as compared to video. Therefore,
higher compression ratio can be achieved in the
range 85%-90%. The PSNR values for texture video
are high which shows that video quality of the
reconstructed video is good.
The proposed algorithm reduces the time
complexity. CUDA allows parallel programming in
CPU and GPU. GPU contains large number of cores
as compared to CPU allowing multiple blocks to run
parallel.
REFERENCES
[1] Zhang Mark Nelson, "The Data Compression
Book: Featuring Fast, Efficient Data
Compression Techniques in C", M&T Books,
1992, ISBN 97815585 12160.
[2] Nvidia GPU Programming Guide GeForce
8 and 9 Series. (developer. download.
nvidia.com)
[3] NVIDIA CUDA Programming Guide,
Version 2.1. (docs.nvidia.com)
[4] K. R. Rao and P. Yip, Discrete Cosine
Transform: Algorithms, Advantages,
Applications.Boston: Academic Press,
1990.
[5] G.Suresh, P.Epsiba, Dr.M.Rajaram, Dr.S.N.
Sivanandam, "Scalable ACC-DCT Based
Video Compression Using Up/Down
Sampling Method", International Journal of
Computer and Network Security, Vol. 2,
No. 6, June 2010
[6] N R Gawande, P P Ghadekar and M L
Dhore,"Efficient Accordion DWT based
Video Coding", International Journal of
Computer Applications 100(10):33-39,
August 2014.
[7] Talukder, Kamrul Hasan, and Koichi
Harada. "Haar wavelet based approach for
image compression and quality assessment
of compressed image." arXiv preprint
arXiv: 1010.4084 (2010).
[8] NVIDIA CUDA C Programming Best
Practice Guide. (docs.nvidia.com)
[9] Stephane Mallat, "A Wavelet Tour of Signal
Processing, Third Edition: The Sparse Way
3rd", Academic Press ©2008, ISBN
9780123743701.
[10] Stanković, Radomir S., and Bogdan J.
Falkowski. "The Haar wavelet transform:
its status and achievements." Computers &
Electrical Engineering 29.1 (2003): 25-44.
[11] P Raviraj and M Y Sanavullah, “The
Modified 2D-Haar WAVELET
Transformation in Image Compression”,
Middle East Journal Scientific Research
2(2),73-78, 2007.
[12] Michael J. Quinn, "Parallel computing (2nd
ed.): theory and practice", McGraw-Hill,
Inc. New York, USA ©1994, ISBN: 0-07-
051294-9.
[13] Péteri, Renaud, Sándor Fazekas, and Mark
J. Huiskes. "DynTex: A comprehensive
database of Dynamic Textures." Pattern
Recognition Letters 31.12 (2010): 1627-
1632.

Dynamic Texture Coding using Modified Haar Wavelet with CUDA

More Related Content

What's hot (20)

Viewers also liked (15)

Similar to Dynamic Texture Coding using Modified Haar Wavelet with CUDA (20)

Recently uploaded (20)

Dynamic Texture Coding using Modified Haar Wavelet with CUDA