Game Development Reference
In-Depth Information
where
P
L
(
)
and
P
˄
(
)
indicate the threshold values from luminance adaptation and
texture masking effect. This model can be further optimized with the local properties
of image content. For example, the local image block can be classified into smooth,
edge and texture types, and for each type, a different compensation factor can be
applied to scale the JND threshold.
Another psychophysical HVS feature-contrast sensitivity function is reflected in
the frequency domain. The contrast sensitivity function describes the visual sensi-
tivity as a function of spatial frequency. Based on the spatial domain JND model,
researchers have developed the frequency domain JNDmodel. The basic JND thresh-
old is first computed by the CSF function, viewing distance, and display device.
Subsequently, with the masking effect within and between each frequency subband,
the final JND is defined as follows: (Lin
2005
),
n
n
JND
SD
(
k
,
b
,
j
)
=
B
k
,
b
,
j
·
M
i
(
k
,
b
,
j
).
(12.2)
i
indicates the JND value in the subband decom-
position with location
k
, subband
b
and frame
j
;
B
In this equation, JND
SD
(
k
,
b
,
j
)
(
,
,
)
represents the JND
computed from the luminance adaptation, texture masking and contrast sensitivity
function; and
M
k
b
j
(
,
,
)
k
b
j
represents the masking effect within and between each
frequency subband.
Moreover, the JND model can be further extended to adapt the properties of input
visual signals. For example, the traditional JND model is only established based
on the luminance signal, but the chrominance component JND model can also be
developed in a similar fashion. In this case, themasking effect within the chrominance
component and between the luminance and chrominance components can also be
exploited to improve the accuracy of the JND model.
For video sequences, it is also observed that the quality perceived by HVS is
not simply reflected by adding or averaging the quality for each frame (Seyler and
Budrikis
1959
; Scharnowski et al.
2007
), and the temporal masking effect can be
taken into account in developing the JND models (Chou and Li
1995
). In Jia et al.
(
2006
), the researchers extended the traditional JND model by considering the eye
moving effect. It is generally believed that, in the temporal domain, fast motion will
decrease the sensitivity of the visual artifacts, while the smoothly moving will cause
the eyes tracking the moving objects. Therefore, different weighting factors can be
employed based upon the traditional JND models. In Chen and Guillemot (
2010
),
based on the visual attention models, the video JND model (FJND) is formulated by
the function of spatial JND (SJND), temporal JND (TJND), and attention model
F
,
FJND
=
f
(
SJND
,
TJND
,
F
).
(12.3)
In the literature, though many JNDmodels have been developed to accurately esti-
mate the visibility threshold, in video coding, the compression artifacts are usually
beyond this threshold and easily perceived by the HVS. In this case, only suprathresh-
old fidelity/distortion measure can be employed to quantify the visual artifacts. The
