Information Technology Reference
In-Depth Information
Quantifying blockiness and packet-loss.
Babu et al. proposed two NR VQA met-
rics - one for blocking and another for packet-loss - in [56]. In essence, a block
is said to be blocky if the edge strength does not have 'enough' variance. Overall
blockiness is then the ratio of blocky blocks to total blocks in a frame. To quantify
packet loss binary edge images are formed using row differences and for each mac-
roblock, the measure of packet loss is the sum of the absolute differences between
the edge images. For each frame, a squared sum is then computed as the final mea-
sure. A comparison of the proposed metrics is undertaken with others proposed in
literature, however this comparison does not involve any subjective correlations.
NR VQA based on artifact measurement.
Farias and Mitra proposed a metric that
is based on measurement of blockiness , bluriness and noisiness [57]. Blockiness is
measured using a modification of Vlachos' algorithm [51]. Width of edges (com-
puted using a Canny edge detector) in a frame is used as a measure of blur. For
each of the 8
3 regions within
a block is computed, and the average of the lowest 4 variances is the block noise
variance. A histogram of these averages is then used to compute a measure of the
frame noisiness. A weighted Minkowski sum of these artifacts is a measure of qual-
ity. Parameters for this sum are estimated using subjective data. Using subjective
data from a small study, the algorithm was shown to perform well.
×
8 blocks, variances of each of the 9 overlapping 3
×
NR VQA based on HVS.
Massidda
et. al.
proposed an NR metric for blur detec-
tion, specifically for 2.5G/3G systems [58]. They computed blockiness, bluriness
and moving artifacts to evaluate quality. Blockiness is evaluated using 8
×
8 non-
overlapping blocks. Within each block they define 4 regions, and sums of horizontal
and vertical edges (obtained using a Sobel filter) are computed over each of these
regions. These values are then collapsed using Minkowski summation to form a sin-
gle value (
B
Sob
). Blur is evaluated using the approach proposed in [59]. Mean and
variance of gradients from two consecutive frames are computed and pooled to ob-
tain a measure of moving artifacts-based distortion. The final quality index is is then
a weighted combination of these three artifacts. The authors evaluate the quality in-
dex as a function of the quantization parameter, instead of using subjective scores.
Prototype NR VQA system.
Dosselmann and Yang propose an algorithm that es-
timates quality by measuring three types of impairments that affect television and
video signals - noise, blocking and bit-error based color impairments [60]. To mea-
sure noise, each frame is partitioned into blocks and inter-frame correlations be-
tween (a subset of) blocks are computed and averaged to form an indicator for that
frame. Using a spatio-temporal region spanning 8 frames, the variance
2
η
σ
is com-
)
p
1
,where
p
1
is experimentally set to
2048. The final noise metric is an average of 256 such values. An alignment based
procedure is used to quantify blocking. Measurement of channel induced color error
is performed by inspecting the R, G and B values from the RGB color space of a
frame and thresholding these values. Even though these measures are not pooled the
2
η
−
−
σ
puted and the noise measure is
η
= 1
(1
