Game Development Reference
In-Depth Information
11.1.2.4 GMSD
It has been found that the pixel-wise gradient magnitude similarity (GMS) between
the reference and distorted images combined with a novel pooling strategy—the
standard deviation of the GMSmap—can predict accurately perceptual image quality
(Xue et al.
2014
).
By convolving horizontal and vertical gradient filters with the reference and dis-
torted images, the corresponding horizontal and vertical gradient maps of
r
and
d
are obtained as follows,
m
r
(
i
)
=
(
r
↗
h
x
)
2
(
i
)
+
(
r
↗
h
y
)
2
(
i
)
m
d
(
i
)
=
(
d
↗
h
x
)
2
(
i
)
+
(
d
↗
h
y
)
)
(11.18)
2
(
i
Then the gradient magnitude similarity (GMS) map is computed via,
2
m
r
im
d
i
+
c
GMS
(
i
)
=
(11.19)
m
r
i
2
+
m
d
i
2
+
c
The pooling strategy is by simply calculating the deviation of the GMS by,
N
1
N
GMSD
=
1
(
GMS
(
i
)
−
GMSM
)
2
(11.20)
i
=
N
i
=
1
GMS
1
where GMSM
=
(
i
)
.
11.1.3 Reduced Reference Image Quality Assessment
The general framework of an RR IQA system is demonstrated in Fig.
11.2
. Taking the
instance of a visual communication system, the features from the original images are
extracted and transmitted to the receiver side. At the receiver, the distance between
the features that are extracted from the original and distorted images is employed in
evaluating the image quality. The bit rate in transmitting the features is typicallymuch
lower than that of the visual information channel. Generally, the selected features
should summarize the whole image properties and exhibit strong correlations with
image distortions, as such the perceptual quality of natural images can be rejected in
the feature comparison process.
Fig. 11.2
General framework of an RR IQA system (Wang and Bovik
2011
)
