Information Technology Reference
In-Depth Information
experimental results, a series of transformations and thresholds are applied to pro-
duce a quality index for the video. A subjective study was carried out and good
correlation with results were demonstrated.
The authors extend the framework proposed in [65] for fluidity measurement in
[69] where a clearness-sharpness metric is introduced to quantify perceptually sig-
nificant blurring in video. The two measures are pooled using a multiplicative term
(in order to account for approximate spatio-temporal separability of the HVS ) and a
final quality metric is produced. A set of videos were used to test the proposed index.
Perceptual Temporal Quality Metric.
Ya n g
et. al.
compute the dropping severity
as a function of the number of frames dropped using timestamp information from
the video stream [70]. They then segment the video temporally into cuts/segments
[71], and determine the motion activity of each such segment using the average
of (thresholded) motion vectors from that segment. The dropping severity is then
mapped onto a perceptually significant dropping factor based on this motion in-
formation. A temporal pooling using a temporal window based approach is then
undertaken. This value is then subjected to a non-linearity and the parameters are
estimated using a fit to subjective data, with hardcoded thresholds. Further weight-
ing and pooling of this temporal indicator leads to the final quality score. Using a
mix of expert and non-expert viewers, a subjective study was undertaken and good
performance on a small set of videos was demonstrated.
NR VQA based on error-concealment effectiveness.
Ya m a d a
et. al.
defined an
error concealment process to be ineffective for a block if the absolute sum of mo-
tion vectors for that block (obtained from motion vector information in the encoded
video stream) is greater than some threshold [72]. Luminance discontinuity is then
computed at the erroneous regions as the mean of absolute differences between
correctly decoded regions and regions where error concealment has been applied.
Another threshold indicates if this region has been concealed effectively. No eval-
uation with respect to subjective perception is carried out, but effectiveness of the
measures are evaluated on a test-set based on packet-loss ratio and number of im-
pairment blocks.
NR modeling of channel distortion.
Naccari
et. al.
proposed a model for channel
induced distortion at the receiver for H.264/AVC [21] coded videos in [49]. The
model seeks to estimate the mean-squared error between the received and transmit-
ted videos - which can also be expressed as the mean distortion induced by all the
macroblocks in a frame. Hence the quantity they wish to estimate is the distortion
induced by the
i
th
macroblock in frame
n
-
D
i
n
. In order to do this, they consider
two cases depending upon whether the macroblock under consideration was lost or
correctly received. In the former case, the predicted distortion is modeled as a sum
of distortions arising from motion vectors, prediction residuals and distortion prop-
agation. For the latter case, the distortion is simply due to error propagation from
the previous frame. The de-blocking filter in H.264/AVC [21] is further modeled
