Big Data-Driven Fast Reducing The Visual Block Artifacts of DCT Compressed Images For Urban Surveillance Systems
Big Data-Driven Fast Reducing The Visual Block Artifacts of DCT Compressed Images For Urban Surveillance Systems
Big Data-Driven Fast Reducing The Visual Block Artifacts of DCT Compressed Images For Urban Surveillance Systems
4, July 2016
ABSTRACT
The Urban Surveillance Systems generate huge amount of video and image data and impose high pressure
onto the recording disks. It is obvious that the research of video is a key point of big data research areas.
Since videos are composed of images, the degree and efficiency of image compression are of great
importance. Although the DCT based JPEG standard are widely used, it encounters insurmountable
problems. For instance, image encoding deficiencies such as block artifacts have to be removed frequently.
In this paper, we propose a new, simple but effective method to fast reduce the visual block artifacts of DCT
compressed images for urban surveillance systems. The simulation results demonstrate that our proposed
method achieves better quality than widely used filters while consuming much less computer CPU
resources.
KEYWORDS
Big Data, JPEG, DCT, Image, Urban Surveillance Systems
1. JPEG INTRODUCTION
The widespread use of video cameras in Urban Surveillance Systems has resulted in an enormous
amount of image and video data, hence the compression of those image and video data become
more and more important in big data research areas. In the mid 80s, the joint collaboration of
International Telecommunication Union (ITU) and International organization for Standard (ISO)
introduced the standard to compress images which is called the Joint Photographic Experts Group
(JPEG) [1]. The core codec method of JPEG is Discrete Cosine Transform (DCT). It may get
high compression ratio and it is easy to be implemented, hence the DCT based JPEG is widely
used [2]. DCT is a Fourier-related transform expresses, it sum up a sequence of cosine functions
at different frequencies.
The two-dimension DCT transform is:
, =
, cos
= 0,1,2 " 1;
DOI : 10.5121/ijdkp.2016.6402
= 0,1,2 $ 1
cos
(1)
19
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
= %
'
=0
1 "1
);
= %
'
=0
1 $1
The two-dimension DCT transform is used in JPEG image compression. And the block diagram
of DCT based JPEG transform can be shown below:
In the coding process, we first break down the image to n*n sub-blocks, normally 8*8 sub-blocks
are chosen. The two-dimension DCT transform is used to get 64 coefficients for every sub-block.
From the two-dimension DCT transform, we can get the 64 coefficients as follows:
From the two-dimension DCT transform, we can get the 64 coefficients as follows:
/.
0
3 cos
3 = %
23
'
5
0
(2)
+=0
)
+>0
Although the application of the DCT formulas would require O(N2) operations, it is possible to
compute the same thing with only O(N log N) complexity. This method is known as fast cosine
transform (FCT) algorithms. In this way, by paying the cost of more additions, we get speed
faster since the CPUs are more excellent to do additions. It makes the DCT become applicable in
real image compression processes. Due to good compression ratio, DCT get efficient memory
utilization and it is widely used in practice.
In fact, the DCT transform does not compress the images, but the quantization process brings
errors into the system. The quantization process is to divide the 64 coefficients by the
quantization matrix and round the results into integers. We know this process will incur errors
into the systems.
Let us assume the quantization matrix is Q[u,v], If the quantization error are expressed as e[u,v],
then the quantitated coefficient matrix are:
7 8*, +9 =
:8.,39
78.,39
+ <8*, +9
(3)
20
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
In order to compress the image, some high frequency information is discarded since human eye
are not sensitive to high frequency information, but this process also incur errors into the systems.
It is obvious that the errors will distribute inside the whole rebuilt image. Every sub-block bring
different errors since they are calculated independently, hence the correlation between different
sub-blocks are destroyed. This leads to the block artifacts effect. For the images, higher
compression ratio leads to more severe block artifacts. Since the block artifacts effect is coming
from the DCT transform, it became the major disadvantage of the JPEG.
In order to show the results of DCT transform, we first simulate the DCT compression effects.
We use an autumn image to be our experiment object. Inside our code, we use different bit rates
to compress the image.
Figure 2 and Figure 3 are both run with the scheme of 8*8 sub-blocks. The compressed images
are more and more blur, and the block artifacts effect shows up more and more severely.
Figure 2 shows the original image and some DCT transformed images with light compression.
Since the qualities of the compressed images are falling down little by little, the details are shown
clearly and the qualities are satisfying for our eyes. The DCT transform appears to be very strong
with light compression.
21
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
Figure 3 shows the DCT transformed images with lower bit rate than that of Figure 2. Inside this
Figure, the block artifacts become more and more obvious, and it is not easy to view the details
inside the image.
2. BACKGROUND
In order to reduce the block artifacts with minimal change of the original coding system, many
researchers have tried different methods. One common method to deal with images is to use
filters to filter the received images. Among the filters, median filtered is a kind of non-linear filter
and is considered suitable to eliminate the random noise [3]. Median filters are widely used in
digital image processing. The idea of the median filter is to run through the signal pixel by pixel,
replacing each pixel with the median value of neighbouring pixels. The pattern of neighbours is
called the "window". For image signals, the windows are normally chosen as 3*3 pixels boxes,
other shapes may be chosen. Another type of widely used filter is adaptive filter [4]. A common
used adaptive filter is wiener filter. Wiener filter is a filter used to produce an estimate of a
desired or target random process by linear filtering of an observed noisy process. It minimizes the
mean square error between the estimated random process and the desired process.
Other methods are used to reduce the block artifacts effect. In [5], the authors proposed a hybrid
method which acts both in the frequency and the spatial domains, to enhance the visual result of
the reconstructed image and reduce the blocking artifacts. In [6], a semi-local approximation
scheme to large-scale Gaussian processed was proposed. This allows learning of task-specific
image enhancements from example images without reducing quality. Similar knowledge-based
algorithms also include simultaneously removed JPEG blocking artifacts and recovers skin
features [7]. On the other hand, some researchers use the property that the original pixel levels in
the same block provide continuity and the correlation between the neighbouring blocks to reduce
22
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
the discontinuity of the pixels across the boundaries [8]. [9] Proposes to remove artificial edges
with a one-time nonlinear smoothing. But the shortcoming for all these methods is that they are
very complex, which will consume a lot of computer CPU and memory resources. Note that the
Urban Surveillance Systems accumulate huge number of images every second, it is impossible to
deal with every image with long time process. It is important to propose very simple but effective
methods to solve the block artifacts effect. The best method should be kept at a single step after
receiving the already affected image. In this way, all the invested hardware and software may be
kept and the cost will be kept low. This motives us to propose the following new simple but
effective method.
23
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
24
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
25
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
Since it is not easy to compare the effects by our eyes, we need some actual numerical results.
Hence, we calculated the Peak Signal to Noise Ratio (PSNR) of all the different methods. PSNR
represents a measure of the peak error; it is a widely used metric to compare image compression
quality. With higher PSNR value, the quality of an image is considered better. Here we show the
PSNR results in Table 1.
Table 1. The PSNR of different situations (Autumn)
Bit rate
0.0938
0.0469
0.0156
Unfiltered
24.9777
23.1011
20.9932
Median filtered
25.0860
23.2140
21.0514
Wiener filtered
25.0572
23.2074
21.1097
Our method
25.2193
23.3213
21.4136
This table shows that, when the images are heavily compressed, the filters and our proposed
method do effect to smooth the sudden changes caused by block artifacts, hence improve the
quality of the images. But our proposed method got better results than these widely used filters.
Our method results in higher PSNRs, which means better qualities.
In order to test our proposed method, we simulate the effects of our proposed method and the two
kinds of widely used filters on another famous image cameraman. The results are shown in
Figures 9, 10 and 11. We also calculated the PSNRs of all the different methods. Here we show
the PSNR results in Table 2.
26
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
27
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
Table 2. The PSNR of different situations (Cameraman)
Bit rate
0.0938
0.0469
0.0156
Unfiltered
23.7897
21.9168
19.4785
Median filtered
23.9410
22.1191
19.5574
Wiener filtered
23.8822
22.0286
19.5837
Our method
23.9571
22.1938
19.8362
Table 2 again shows that our proposed method results in better quality. The main benefit of our
proposed method is that the comparably better PSNR results are gotten with very simple
calculations. Considering that normal filters need to calculate every pixel for an M*N image
while our proposed method only need to calculate the pixels at the block borders, our method
only need a small fraction of the CPU calculation source comparing with a normal image filter.
4. CONCLUSIONS
The Urban Surveillance Systems are facing the problem of huge amount of video data, and the
compression technique is the most fundamental and important to deal with the data. Although the
DCT based JPEG standard are widely used, it encounters problems such as block artifacts. In this
paper, in order to reduce the block artifacts, we proposed a very simple but effective method. The
simulation results show that our proposed method results in better quality than widely used filters
while consuming much less computer CPU resources. In this way, we could reduce the visual
block artifacts of DCT compressed images fast for Urban Surveillance Systems.
ACKNOWLEDGEMENTS
This work was supported by the Lancaster China Smart City Urban Surveillance Systems project
under Grant number CSA7035.
REFERENCES
[1]
R. Singh and V.K. Srivastava, JPEG2000: A review and its performance comparison with JPEG.
Power, Control and Embedded Systems (ICPCES), 2012 2nd International Conference on, 2012,
Pages: 1-7.
[2]
M. Jha, A.S.Yumnam, D. Raju, Comparison between image codecs: JPEG and JPEG2000,
Computing for Sustainable Global Development (INDIACom), 2014 International Conference on,
2014, Pages: 532 - 535.
[3]
S. Vishaga and S. L. Das, A survey on switching median filters for impulse noise removal, Power
and Computing Technologies (ICCPCT), 2015 International Conference on Circuit, 2015, Pages: 1 6.
[4]
A. N. Venetsanopoulos and K. N. Plataniotis, Adaptive filters for color image processing: A survey,
Signal Processing Conference, 2000 10th European, 2000, Pages: 1 - 4.
[5]
[6]
Y. Kwon, K. I. Kim, J. Tompkin, J. H. Kim, C. Theobalt, Efficient Learning of Image SuperResolution and Compression Artifact Removal with Semi-Local Gaussian Processes, IEEE
Transactions on Pattern Analysis and Machine Intelligence, 2015, Volume: 37, Issue: 9, Pages: 1792 1805.
28
International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.4, July 2016
[7]
[8]
Y. Luo, R. K. Ward, Removing the blocking artifacts of block based DCT compressed images,
IEEE Transactions on Image Processing, 2003, Volume: 12, Issue: 7, Pages: 838 - 842.
[9]
J. Chou, M. Crouse, K. Ramchandran, A simple algorithm for removing blocking artifacts in blocktransform coded images, IEEE Signal Processing Letters, 1998, Volume: 5, Issue: 2, Pages: 33 - 35.
29