Information Technology Reference
In-Depth Information
using attenuation coefficients. Correlation between estimated channel distortion and
measured channel distortion is estimated for performance evaluation.
NR VQA for HD video.
Keimel
et. al.
proposed an NR VQA algorithm specifi-
cally for HD video [73]. Features extracted for this purpose include blur, blockiness,
activity and predictability using previously proposed methods. Different pooling
strategies are used for each of these features, followed by principal component anal-
ysis (PCA) for dimensionality reduction and partial least squares regression (PLSR)
to map the features to visual quality. A correction factor is then incorporated based
on a low-quality version of the received video to find the final quality score. The
astute reader would have observed that the proposed method is pretty general and
does not specifically address HD video. The proposed algorithm is tested on HD
video however and decent performance is demonstrated.
Video quality monitoring of streamed videos.
Ong
et. al.
model jerkiness between
frames using absolute difference between adjacent frames and a threshold [74]. Pic-
ture loss is similarly detected with another threshold. Blockiness is detected using
using a technique similar to those proposed previously. The test methodology is
non-standard and requires users to identify number of picture freezes, blocks and
picture losses in the videos. Perceptual quality is not evaluated however.
Other techniques for NR VQA include the one proposed in [75] which we do not
explore since a patent on the idea was filed by the authors. Further, many of the met-
rics discussed here were evaluated for their performance using a variety of criteria in
[76, 77]. Hands
et. al.
provide an overview of NR VQA techniques and their appli-
cation in Quality of Service (QoS) [78]. Kanumri
et. al.
model packet loss visibility
in MPEG-2 video in [79] to assess quality of video.
6Conluion
In this chapter we began with a discussion of video quality assessment and intro-
duced datasets to evaluate performance of algorithms. We then went on to describe
the human visual system briefly. A summary of recent reduced and no reference
algorithms for quality assessment then followed.
We hope that by now the reader would have inferred that NR VQA is a dif-
ficult problem to solve. It should also be amply clear that even though a host of
methods have been proposed (most of which are listed here) there does not seem to
emerge an obvious winner. Our arguments on the use of a common publicly avail-
able dataset for performance evaluation are hence of importance. The reader would
have observed that most authors tend to select a particular kind of distortion that
affects videos and evaluate quality. Any naive viewer of videos will testify to the
fact that distortions in videos are not singular. In fact, compression - which is gen-
erally assumed to have a blocking distortion, also introduces blurring and motion-
compensation mismatches, mosquito noise, ringing and so on [80]. Given that there
exist a host of distortions that may affect a video, one should question the virtue of
