Information Technology Reference
In-Depth Information
RR Video Structural Similarity.
The Structural SIMilarity index (SSIM) was pro-
posed as a FR image quality assessment technique in [46] and extended to VQA in
[47]. Albonico
et. al.
proposed a RR VQA in [48] where the extracted features were
similar to those proposed for video SSIM. Using 16
×
16 blocks the mean
μ
x
(
i
,
n
)
x
(
i
,
n
) for the
i
th
macroblock in the
n
th
frame is computed
at the source and are transmitted as the RR feature vector. At the receiver end, a dis-
tortion estimation is undertaken using the technique described in [49] and reviewed
in the next section to produce
D
(
i
,
n
). Further, the mean
and standard deviation
σ
μ
x
(
i
,
n
) and standard devia-
tion
σ
x
(
i
,
n
) are estimated at the receiver end from the received frame. A covariance
estimate is then formed as:
σ
x
(
i
,
n
)
2
+
σ
x
(
i
,
n
)
2
+(
−
μ
x
(
i
,
n
))
2
−
D
(
i
,
n
)]
σ
xx
(
i
,
n
)=0
.
5
×
[
μ
x
(
i
,
n
)
This allows for computation of the SSIM index for each frame. The authors state
that the index estimated in such a fashion fails to match-up to the original SSIM
value and hence a lookup-table based approach is used to eliminate a
bias
in the
estimated values. The authors claim that this allows for an estimate of the SSIM
index and hence of video quality.
5
No-Reference Algorithms
No-reference algorithms are those that seek to assess quality of a received video
without any knowledge of the original source video. In a general setting, these al-
gorithms assume that the distorting medium is known - for example, compression,
loss induced due to noisy channel etc. Based on this assumption, distortions specific
to the medium are modeled and quality is assessed. By far the most popular distort-
ing medium is compression and blockiness and bluriness are generally evaluated for
this purpose. We classify NR algorithms as those based on measuring blockiness,
those that seek to model the effect of multiple artifacts (for example, blocking and
blurring ) and other techniques based on modeling the channel or the HVS .
5.1
Blockiness-Based Techniques
Blockiness measure for MPEG-2 Video.
Tan and Ghanbari [50] used a harmonic
analysis technique for NR VQA which was also used in [42] (for RR VQA). A So-
bel operator is used to produce a gradient image, which is then subjected to a block
FFT . The ratio of sum of harmonics to sum of all AC components within a block is
computed in both horizontal and vertical directions. The phase of harmonics across
the frame are then histogrammed. The authors suggest that a smaller standard devi-
ation in the (empirical) probability density function (PDF) indicates greater block-
iness. Based on certain pre-set thresholds on the harmonic ratio of the magnitudes
and the phase of the harmonics a block is considered to be 'blocky'. The authors
however test their technique on I and P-frames and state that the method does not
function well for B-frames.
