Emna CHEBBI et al.

, International Journal of Advanced Trends in Computer Science and Engineering, 1 (2), May June, 2012, 50 - 55
50
@ 2012, IJATCSE All Rights Reserved






Image Quality Assessment based on Perceptual Blur Metric


Emna Chebbi, Faouzi Benzarti, Hamid Amiri
Signal, Image Processing and Patterns Recognition (TSIRF) Laboratory
National Engineering School of Tunis (ENIT), Tunisia
emna.chebbi@gmail.com, benzartif@yahoo.fr, hamdilamiri@yhaoo.com


ABSTRACT

The recent development of digital image acquisition
technologies leads to better image quality, in terms of spatial
resolution and sensitivity. Image quality is a characteristic of
an image that measures the perceived image degradation.
Several techniques and metrics are proposed which can be
classified as Full-Reference (FR) method, No-Reference (NR)
method and Reduced Reference (RR) method. In this field of
image quality assessment, it is crucial to deep research the
physiology and psychology of human visual system. However,
it is obvious that strong correlation between the results and
human visual perception is essential. In this paper, we propose
a new approach for image quality assessment that combines
the perceptual blur metric and the index of Structural
Similarity (SSIM) in order to improve the image quality
quantification.

Keywords : Image Quality, Image Assessment, Human Visual
System, Structural Similarity, Blur perception

1. INTRODUCTION

Visual images are the most important and data intensive means
for humans to acquire information, digital image acquisition,
communication, storage processing, and display devices have
become ubiquitous in daily life. Since digital images are
subject to a wide variety of distortions in any of these, and
since image traffic has become quite dense, the assessment of
digital image quality has become an exceedingly important
topic. In fact, the lack of information caused by the processing
is results in an alteration of the original image, it is
necessary to evaluate the loss of image quality by
comparing the distorted image to the original one. The first
approach involves using human observers to assess image
quality. However, this method has several drawbacks. First,
it uses the subjectivity of human beings which assess
the image quality in different ways. In addition, it provides a
qualitative result, as we would like to have a quantitative
result. Finally, this subjective assessment is tedious and is
difficult to apply when the number of images to be treated is
important.



Therefore, objective methods for image quality assessment
have been proposed. These metrics should provide quality
scores consistent with human judgment which requires
the integration of the main properties of the Human Visual
System (HVS). The objective assessment of image quality or
video is based on many criteria for determining an objective
quality score. These criteria are classified into three categories
according to the information necessary for the assessment: Full
Reference (FR) [4][5], Reduced Reference (RR) [16][17] and
No Reference (NR) [2][13]. The FR methods require the
disposal of the reference image and the degraded version to
assess. These approaches are used in introducing degradation
systems, like systems of loss compression which aims to
estimate the amount of distortion caused by the compression
and quality of the resulting image. Generally, these
approaches are based on modeling the Human Visual System
(HVS) which they incorporate one or more properties of this
model. The RR methods provide a measure of quality with
only a small set of features measured on the
reference. However, the use of a limited amount of
information to develop a final quality score is much harder
than full reference methods. They are used in a transmission
where it is impossible to transmit all information related to
both versions of the image; the reduced reference is then
encoded and transmitted with the reference version to judge
the quality. The NR methods assess the quality of an image
without referring to the reference image. These methods were
the most difficult to develop since they are based on claims
that the image should be. These approaches are popular
because they do not require the transmission of the reference
version to evaluate the quality of the transmitted image.

This paper is organized as follows: In the next section, we
briefly review some related works in the image quality
assessment metrics. In section 3, we present the blur detection
using discrete wavelet transform. The proposed method is
presented in section 4. Experimental results are presented in
section 5. Finally conclusion is given is section 6.


ISSN No. 2278 -3091
Volume 1, No.2, May June 2012
International Journal of Advanced Trends in Computer Science and Engineering
Available Online at www.warse.org/ijatcse/info.html

Emna CHEBBI et al., International Journal of Advanced Trends in Computer Science and Engineering, 1 (2), May June, 2012, 50 - 55
51
@ 2012, IJATCSE All Rights Reserved


2. RELATED WORKS

In this section we briefly present some related works in the
image quality assessment metrics; which are divided into two
main classes as shown the Figure 1.


Figure 1: Image quality measures
The objective measure most commonly used is that of Index of
Structural Similarity (SSIM) [19]. This metric is an enhanced
version of UQI [18]. The index UQI is easy to implement and
can be applied to different types of images. It measures the
uncorrelation between the reference image and the distorted
image, as well as the degradation of the luminance component
and the contrast between the two versions of the image. This
criterion is then determined by the product of these three
measures for comparing the luminance l( x, y) , the contrast
c( x, y) and structure s( x, y) between two signals x and y.

l( x, y) =
2
x

2
+
2
, c( x, y) =
2c
x
c
j
c
x
2
+c
j
2
, (1)
s( x, y) =
co:
x
o
x
o

Where p
x
means the average of x, p

average of y, o
x
2
the
variance of x , o

2
variance of y and co:
x
the covariance
between x and y.
SSIM is obtained in case where p
x
2
+ p

2
or o
x
2
+ o

2
have
values close to zero. The formula is then given by:

SSIH( x, y) =
(2
x

j
+c
1
)(2co
xj
+c
2
)
(
x
2
+
j
2
+c
1
)(c
x
2
+c
j
2
+c
2
)
(2)

wherec
1
= ( k
1
I)
2
, c
2
= ( k
2
I)
2
, wit I refers to the dynamic
values of the pixels, or 255 for images coded on 8 bits, by
default k
1
= 0.01 and k
2
= 0.03 .

Among the most recent objective measures incorporating the
human visual system is the Visual Information Fidelity (VIF).
The VIF metric [9] is the result of an improvement in the
Information Fidelity Criterion [10] which is to integrate a
normalization step. This metric is obtained by the relation:

IIF =
I(C

N,]
;P
N,]
\ s
N,]
)
]
I(C
N,]
;L
N,]
\ s
N,]
)
]
(3)

where I(C

N,]
; F

N,]
\ s
N,]
) and I(C

N,]
; E

N,]
\ s
N,]
) refer to
the information extracted from the original image and its
degraded version from the sub-band j.

Other objective measure that integrates the SVH is that of
Visual Signal to Noise Rate (VSNR) [6]. The VSNR metric is
a quality measure based on analysis of wavelet coefficients; it
is divided into two steps. The first is to ensure that degradation
is well above the threshold of visibility before measuring. This
check is performed in each sub-band of wavelet
decomposition. The second step is to evaluate the perception
of degradation above the threshold of visibility.
VSNR is determined by:

ISNR = 10log[
C
2
( x)
P
2
(4)

Where C(x) denotes the average contrast of the reference
image and DP is the perceptual distortion. This metric does not
include color management or the spatial location of damage.
However, measurements of distances and structural
approaches are still consistent with the trial staff giving birth
metrics based on the integration of the properties of HVS.
Moreover, global quality measures have been defined recently
such as the Neural Fusion Approach [1]. Its is a new metric
with full reference based on the fusion of several conventional
metrics with full reference using the learning algorithm
artificial neural networks. The fusion brings performance more
important than using individual metrics. Indeed, the existing
metrics with full reference do not always produce excellent
results for all types of degradation. To confront this limitation,
it is appropriate to use a process of ranking and then use an
artificial neural network. It is to combine the best performing
metrics in a single metric called Index of Global Quality. The
image database used TID 2008 shows that this Index provides
assessment image quality results consistent with the subjective
assessment.

3. PERCEPTUAL BLUR DETECTION

A blur detection scheme is proposed using Haar wavelet
transform [11][14]. It may not only judge whether or not a
particular image is blurred but also to identify how the data
image is blurred which is based on edge sharpness
analysis. The proposed scheme benefits from the ability
of Haar wavelet transform in both discriminating various types
of edges and sharp recovery of the blurred version. In fact,
Wavelet transform is well known for its ability to analyze
multiple resolutions. Based on an important fact that the local
maxima of a wavelet transform to detect the localization of
irregular structures is proved [2][12].


Image qual i t y
assessment
Subj ect i ve
assessment
M ean Opi ni on Scor e
(M OS)
Just Not i ceabl e Di f f er ence
(JND),...
Obj ect i ve
assessment
Ful l Ref er ence
Di st ance measur es (M ES, PSNR, SNR, ...)
M et r i cs model i ng t he human vi sual
syst em (UQI, SSIM , M SSIM , VSNR, WSNR,
PSNR-HVS, IFC, VIF, ...)
Reduced
Ref er ence
Zi pf l aw
Car nec met r i c, ...
No Ref er ence
Bl i nd i mage qual i t y i ndex (BIQI)
Jpeg qual i t y scor e
Emna CHEBBI et al., International Journal of Advanced Trends in Computer Science and Engineering, 1 (2), May June, 2012, 50 - 55
52
@ 2012, IJATCSE All Rights Reserved


L L 3 H L 3 H L 2 H L 1: H or i zontal Detai l
(l evel 1)
L H 3 H H 3
L H 2 H H 2
L H 1: Ver ti cal Detai l

H H 1: Di agonal Detai l
(l evel 1)
Figure 2: Haar wavelet decomposition at level 3

The procedure to be followed to detect edge using Haar
wavelet transforms [8][15] starts by a Haar wavelet
decomposition level 3 of the distorted image (Figure 2).
Afterwards, it comes to build the edge map for each level by
the following formula:

E
mup
i
=
IE

2
+ EI

2
+ EE

2
; i = 1,2,3 (5)

The next stage consists in partitioning the edge maps and
determines the local maximum for each window. For the
highest level, the window size is 2x2. The low level is 4x4 and
the lowest is 8x8. Eventually, for each edge map, the edges
blurred must be determined. The Haar wavelet transform has
the benefit to recover the thin edges blurred which leads to
determine the number of edge blurred and so to extent the blur
amount. The Haar wavelet transform ducts to different rules
applied to detect the blur in the image presented. For each
edge point, for a given threshold, it is to identify both of edges
points which are more likely to be in the blurred image and
edge points that lost their intense.

In fact, if Emax
1
(k,l)>threshold or Emax
2
(k,l)>threshold or
Emax
3
(k,l)>threshold then (k,l) is an edge point where Emax
i
is the local maximum for the level i. For each edge point, it
comes to identify the type of edge. Most natural image
contains all types of edges: Dirac-structure, Roof-structure and
step-structure which is divided into Astep-structure and Gstep-
structure.

For each edge point (k,l):
if Emax
1
(k,l)>Emax
2
(k,l)>Emax
3
(k,l) then (k,l) is a
Dirac-structure or Gstep-structure,
if Emax
1
(k,l)<Emax
2
(k,l)<Emax
3
(k,l) then (k,l) is a
Roof-structure or Gstep-structure,
if Emax
2
(k,l)>Emax
1
(k,l) and Emax
2
(k,l)>Emax
3
(k,l)
then (k,l) is Roof-structure.
For each edge point (k,l) Gstep-structure or Roof-structure, if
Emax
1
(k,l)<threshold then (k,l) is more likely to be in a
blurred image.

Based on these Haar wavelet rules, an image is judged as
blurred if the ratio between Dirac-Structure and Roof-Structure
is superior of 0.05. Once there is presence of blur, it consists
then to determinate the blur amount existing in the distorted
image by calculating the ration between the numbers of edges
blurred Gstep-Structure and Roof-Structure.


4. PROPOSED APPROACH

The proposed method as shown in Figure 3, is to merge
the quality score obtained by the index of structural similarity
with the blur amount measured in order to obtain a quality
measure that takes account the perception of blur. The most
important assumption is that the human eye is typically
suitable for the extraction of structural information of an
image. It is then necessary to measure the degradation of
this structural information. The idea is to extract local
structural attributes of the image from which each block is
described by its brightness, contrast and structure.

We start the image quality assessment by using SSIM without
introducing the blur factor. The results provided by SSIM are
then compared to those delivered over the new metric
proposed. The purpose is to develop a tool to improve the
quantification of image quality. To achieve this aim, we
exploit the objective methods based on measuring perceptual
quality of an image. These methods consist in measuring the
error between a visibility degraded image and a reference
image using a variety of known properties of the visual system
Human (HVS). By exploiting the concept of HSV to which the
human visual is highly suitable for extract the structural
information of an image. We have exploited this concept in
quality assessment image. This is reflected in the measurement
of structural information as an index structural similarity
SSIM. However, SSIM doesnt take account of the blur
detection factor and to its extent [4].




















Figure 3: Flow chart of the Proposed Approach

5. EXPERIMENTAL RESULTS

We choose for our experimental results the CSIQ database [7].
It consists of 30 original images; each is distorted using six
different types of distortions at four to five different levels of
Or i gi nal image Degr aded i mage
I mage qual i ty
assessment by
SSI M
I M
Bl ur amount
Scor e fusion
Fi nal
Assessment
Emna CHEBBI et al., International Journal of Advanced Trends in Computer Science and Engineering, 1 (2), May June, 2012, 50 - 55
53
@ 2012, IJATCSE All Rights Reserved


distortion. CSIQ images are subjectively rated base on a linear
displacement of the images across four calibrated LCD
monitors placed side by side with equal viewing distance to
the observer. Each original image in the database is distorted
using six different types of distortions at four to five different
levels of distortion. The distortions used in CSIQ are: JPEG
compression, JPEG-2000 compression, global contrast
decrements, additive pink Gaussian noise, and Gaussian
blurring the purpose of our research. In our experiments, we
used a few set original images presented by the Figure 4.


Figure 4: The original images of CSIQ data base


Figure 5: Blur images

The Figure 6 shows the mean of SSIM values versus blur
standard deviation of a set of blurred images (Figure 5).
For each blur image, it comes to determinate the score of
SSIM for a number of iteration well defined.

Figure 6: Mean of SSIM

The mean of SSIM values is represented by a decreasing curve
(Figure 6), a rise in the value of blur standard deviation
resulted in a decline in score of SSIM.
The fitted curve is a cubic polynomial:

ssim( x) = p1 x^ 3 + p2 x^ 2 + p3 x + p4 ( 6)

where the coefficient of the polynomial (with 95% confidence
bounds):
p
1
= 3.89c 005; p
2
= 0.001516 ; p
3
= 0.008339;
p
4
= 0 .8899.

According to results, SSIM is less sensitive to the change of
blur amount in an image. So, it becomes vital to determinate
the blur extent of an image. The Figure 7 presents the mean of
blur amount already calculated for all degraded images.


Figure 7: Mean of blur amount
The fitted curve of Blur amount is determined by the following
formula:

blur( x) = o1 sin( b1 x + c1 ) + o2 sin( b2 x +
c2) + o3 sin( b3 x + c3) ( 7)

Emna CHEBBI et al., International Journal of Advanced Trends in Computer Science and Engineering, 1 (2), May June, 2012, 50 - 55
54
@ 2012, IJATCSE All Rights Reserved


Where the coefficients are determined with 99% confidence
bounds:
o1 = 1.388 ; b1 = 0.129 ; c1 = 0.04447 ;
o2 = 0.7787; b2 = 0.2232 ; c2 = 2.307;
o3 = 0.1553; b3 = 0.3624; c3 = 3.842

It should then apply the proposed formula to get the final score
of a blur image:

SSI M_blur = SSIM_Score-(Blur Extent*) (8)

where =10. If BlurExtent=0, the image is judged unblurred
and the quality measure takes the value of SSIM score.


Figure 8: Mean of SSIM_blur

The proposed approach is presented by the figure below
(Figure 8); this is a downward curve. An increase in the value
of the blur standard deviation leads to a lower score of the
proposed quality measure. According to experiments, the
scores provided by the proposed approach are more declined
than SSIM; which reflects its effectiveness in assessing quality
of blurred images. The fitted curve of the proposed approach
is a cubic polynomial which is determined by the following
formula:

ssim
bIu
( x) = p1 x^ 3 + p2 x^ 2 + p3 x + p4 (9)

where the coefficients are determined with 99% confidence
bounds:
p1 = 5.048c 005 ; p2 = 0.001869;
p3 = 0.009232; p4 = 0 .8323

As it shown the results, the integration of the blur factor in
measuring the structural metric SSIM improves scores
obtained for the assessment measure of image quality and
provides satisfactory and relevant results that are consistent
with the human eye.




6. CONCLUSION

In this paper, a new approach for image quality assessment has
been proposed based on blur estimation. Experimental results
indicate that our metric significantly exceeds the performance
of the SSIM and provided results also correlate with the
human visual system. These results are encouraging as an
approach to conception of a metric of image quality taking into
account the perception of blur. They confirm the relevance
of the develop approach by incorporating a model to
aim psycho visual the aspect in this calculation.

REFERENCES

1. A.Chetouani, A.Beghadadi et and M.Deriche. A universal
full reference image quality metric based on a neural
fusion approach, Proceedings of 2010 IEEE 17
th

International Conference on Image Processing.

2. A. Ciancio, A.L.T da Costa, E.A.B da Silva, A. Said, R.
Smadani and P. Obrador. No-Reference Blur Assessment
of Digital Pictures Based on Multifeature Classifiers,
IEEE Transactions on Image Processing, Vol. 20, No. 1,
January 2011.

3. A. K. Moorthy and A. C. Bovik. A Modular Framework
for Constructing Blind Universal Quality Indices,
submitted to IEEE Signal Processing Letters, 2009.

4. C.Li et and A.C.Bovik. Content partioned structural
similarity index for image quality asssesment, Image
communication, 2010.

5. C.Yim and A.C.Bovik. Quality Assessement of
Deblocked Images, IEEE Transaction on Image
Processing, 2011.

6. D. Chandler and S. Hemami. Subjective image database,
http://foulard.ece.cornell.edu/dmc27/vsnr/vsnr.html,
2007.

7. E. C. Larson and D. M. Chandler. Most apparent
distortion: full-reference image quality assessment and
the role of strategy, Journal of Electronic Imaging, 19
(1), March 2010.

8. G. Cao, Y. Zhao and R. Ni. Edge-based Blur Metric for
Tamper Detection, Journal of In formation Hiding and
Mutimedia Signal Processing, Ubiquitous International,
Vol. 1, No. 1, 2010.

9. H.R. Sheikh and A.C Bovik. Image information and
visual quality, IEEE Transactions on Image Processing,
Vol. 15, pp. 430-444, 2006.



Emna CHEBBI et al., International Journal of Advanced Trends in Computer Science and Engineering, 1 (2), May June, 2012, 50 - 55
55
@ 2012, IJATCSE All Rights Reserved


10. H.R. Sheikh, A.C. Bovik and G. de Veciana. An
information fidelity criterion for image quality
assessment using natural scene statistics, IEEE
Transactions on Image Processing,Vol. 14, pp.2117-2128,
2005.

11. H. Tong, M. Li, H. Zhang and C. Zhang. Blur
Detection for Digital Images Using Wavelet
Transform, Multimedia and Expo, 2004, ICME '04.
2004 IEEE International Conference.

12. I. Makaremi and M. Ahmadi. Blur invariants : A novel
representation in the wavelet domain, Departement of
Electrical and Computer Engineering University of
Windsor, ON, Canada N9B 3P4, 2010.

13. L. Liang, S. Wang, J. Chen, S. Ma, D. Zhao et and
W. GAo. No-reference perceptual image quality metric
using gradient profiles for JPEG2000, Image
Communication, 2010.

14. R. Liu, Z. Li and J. Jia. Image Partial Blur Detection
and Classification, Computer Vision and Pattern
Recognition, 2008, CVPR 2008, IEEE Conference on
Hong Kong.

15. S. Wu, W. Lin, S. Xie, Z. Lu, E.P. Ong and S. Yao. Blind
blur assessment for vision-based applications, J. Vis.
Commun. Image R. 20 (2009) , pp.231-241.

16. U. Engelke, M. Kusuma, H-J. Zepernick and M. Caldera.
Reduced-reference metric design for objective
perceptual quality assessment in wireless imaging,
Signal processing: Image communication, 2009.

17. Y. Caron, P. Markis et and N. Vincent. Compressed
Image Quality Evaluation using Power Law Models,
IEEE 2006.

18. Z. Wang, A.C. Bovik, H.R. Sheikh and E.P. Simoncelli.
Image Quality Assessment: From Error Meaurement
to structural similarity, IEEE Transactions on Image
Processing, Vol. 13, January 2004.

19. Z. Wang, E.P. Simoncelli and A.C. Bovik. Multi-scale
structural similarity for image quality assessment,
Asilomar Conference on Signals, Systems and Computers.