![](bg2.png)
sults have shown that our approach is able to achieve
high accurate image quality rankings using an objec-
tive evaluation metric.
2 QUALITY ASSESSMENT
Given an ideal black and white image with intensity
values µ
1
and µ
2
, the histogram consists of only two
sharp peaks at these two intensity values. For a real
scene image captured by a camera, the intensities are
spread out due to the optical system and noise, and
the histogram usually contains two bell-shape distri-
butions located at µ
1
and µ
2
. If the image contains
defocus blur, the mixture of high and low intensity
values introduced by the point spread function gen-
erates a smooth transition between µ
1
and µ
2
in the
histogram.
According to the defocused image formation, the
histogram changes with the blur extent (Lin et al.,
2012). When the defocus blur becomes severe, the
two main lobes corresponding to the high and low in-
tensity regions diminish, and the transition area be-
tween the two main lobes increases. Thus, the blur
extent of a defocused image can be characterized by
the distribution of its histogram. By comparing the
histogram of the unknown defocused image with the
histogram of a calibrated image, the blur extent of the
unknown image can be identified. More specifically,
the blur parameter of the point-spread function can
be derived by this histogram matching technique and
used to represent the amount of defocus blur associ-
ated with the given image (Lin and Chou, 2012).
To apply our blur identification technique for im-
age sharpness evaluation, we need to select several re-
gions of interest (ROI) for histogram matching. This
is accomplished automatically by performing the fol-
lowing steps. First, an edge image obtained from
Canny edge detection is used to derive suitable edge
segments for blur extent estimation. Since the blur
identification is carried out locally along the horizon-
tal direction, the edge segments are constrained by
three criteria to ensure the robustness of histogram
matching: (a) the vertical 8-neighbor connectivity, (b)
a minimum edge length threshold (typically about 1%
of the original image height), and (c) no other edges
present in the neighborhood.
Second, an initial ROI with a fixed width (typi-
cally about 2% of the original image width) is as-
signed for each edge segment. The intensity distri-
bution of each ROI is analyzed, and only those ROIs
with low intensity variation on both sides of the edge
segment are preserved. Finally, each ROI is enlarged
in the horizontal direction if the local intensity distri-
butions on both sides of the edge still remain uniform
when including an extra column of pixels from the
left and right of the ROI respectively. This process
is carried out iteratively until the local intensity vari-
ation is no longer uniform. It aims to provide larger
ROIs for histogram matching and achieve better blur
identification results.
After the ROIs are selected for a given image, his-
togram matching is performed on each ROI individu-
ally. The average of the identified blur extents from all
ROIs is used to represent the image sharpness. For a
given set of images, the quality ranking is then derived
based on the amount of their blur extents. To evalu-
ate the performance of our image quality assessment
technique, the ground truth image quality ranking is
used for comparison. Suppose a set of n images is
indexed by 1, 2, · · · , n, according to their ground truth
quality, and the evaluated quality ranking is given by
a permutation function p(·). Then the quality assess-
ment score for the image set is defined by
S =
∑
n
i=1
∑
n
j=1
c(i, j)
C
n
2
(1)
where
c(i, j) =
1 if i < j implies p(i) < p( j)
0 otherwise
(2)
It is seen that the quality assessment score S ∈ [0, 1].
The special cases S = 1 and S = 0 correspond to the
correct and completely reverse quality rankings, re-
spectively.
3 EXPERIMENTAL RESULTS
The proposed image quality assessment technique has
been tested using the images with synthetic and real
defocus blur. For the experiments with synthetic blur,
we choose 8 sets of images with Gaussian blur from
the LIVE image database (Sheikh et al., 2006; Wang
et al., 2004). A series of 10 blurred images are gen-
erated from each reference image using the circular-
symmetric 2-D Gaussian kernels on with standard de-
viation ranging from 0.5 to 5 pixels with mask size:
5,9,13,17,21,25,31,35,39,43.
In the experiments, the number of ROIs extracted
from each test image and used for image quality as-
sessment ranges from 21 to 107. The blur identifi-
cation results of the LIVE database images are tabu-
lated in Table 1. The quality evaluation of the image
datasets is illustrated in Figure 1. Index number 0
indicates the reference or focused image, and the im-
ages generated with more severe blur are those with
VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications
102