Information Technology Reference
In-Depth Information
DCT domain before transmission. At the receiver side, a similar process is under-
taken to estimate the same parameters on the degraded video. The KL divergence
between the original parameters and the distorted parameters is computed and av-
eraged across subbands to form a distortion measure for the video. The proposed
algorithm is not tested on a public dataset, but instead a small set of videos are used
for evaluation.
Representative-Luminance based RR VQA.
The essence of the idea proposed by
Ya m a d a
et. al.
in [41] is to estimate PSNR at the receiver using luminance infor-
mation embedded in the transmitted video stream. Block variance of each 16
16
block is evaluated and the representative luminance of a frame is chosen from a
subset of the blocks which have variance equal to the median variance of the frame.
The authors claim that this captures the luminance of pixels in the medium fre-
quency range. PSNR is computed at the receiver using this additional information
and is used as the quality metric.
×
RR VQA based on Local Harmonic Strength.
Gunawan and Ghanbari proposed
a RR VQA algorithm in [42] based on local harmonic strength. First a Sobel filter is
used to produce a gradient image [43]. This image is then segmented into blocks and
a harmonic analysis is applied on each of these blocks. Harmonic analysis consists
of applying the 2-D fast Fourier transform (FFT) [44] on a block-by-block basis
and computing the magnitude of the transform at each pixel location. The local har-
monic strength is the sum of the magnitudes of the transform at particular locations
within a block. The local harmonic feature is used as the RR feature. A similar
analysis is performed at the receiver on the distorted video and harmonic gains and
losses are computed as differences between the harmonic features of the reference
and distorted videos. A motion correction factor obtained from the mean of motion
vectors (computed using a block-based motion estimation algorithm) is then applied
to obtain the corrected harmonic gain/loss. The quality measure of the sequence is
a linear combination of these corrected harmonic features. The parameters of the
combination are obtained using a small training set.
Distributed Source Coding based estimation of channel-induced distortion.
In
[45], each macroblock in a frame is rasterized (i.e., converted into a vector)
x
(
k
)
and
then a RR feature vector
y
is computed, where each entry of the feature vector is
y
i
=
a
T
x
(
k
)
where
a
is s pseudo random vector with
= 1. This vector
y
is then subjected to
a Wyner-Ziv encoding [40], in order to reduce the bit-rate. At the receiver, a similar
process is carried out using the transmitted pseudo random vector and an estimate of
the mean square error (MSE) is obtained between the transmitted RR feature vector
and the one computed at the receiver. The authors claim that the method estimates
MSE well with a small increase in transmission bit-rate (for the RR features).
||
a
||
