Information Technology Reference
In-Depth Information
authors demonstrate how these measures can be incorporated into a system. Corre-
lation with subjective assessment is not studied.
NR VQA based on frame quality measure.
Kawayoke and Horita proposed a
model for NR VQA consisting of frame quality measure and correction, asymmet-
ric tracking and mean value filtering [61]. The frame quality measure is simply
neighboring pixel differences with and without edge preservation filtering. This is
histogrammed, pooled and corrected to obtain the final measure. Asymmetric track-
ing accounts for the fact that humans tend to perceive poorer regions with greater
severity than good ones [62]. Finally mean value filtering removes high frequency
ingredients from the measure to produce the quality index. Evaluation on a small
dataset shows good correlation with perception.
5.3
Other Techniques
NR VQA incorporating motion information.
Ya n g
et. al.
proposed a NR VQA
which incorporates motion information in [63]. Block-based motion estimation is
applied on a low-pass version of the distorted video. Translation regions of high
spatial complexity are identified using thresholds on variance of motion vectors and
luminance values. Using the computed motion vectors for these regions, sum of
squared error is computed between the block under consideration and its motion
compensated block in the previous frame, which is then low-pass filtered to give a
spatial distortion measure. A temporal measure is computed using a function of the
mean of the motion vectors. A part of the VQEG dataset is used to train the algo-
rithm in order to set the thresholds and parameters in the functions. Testing on the
rest of the dataset, the algorithm is shown to correlate well with human perception
of quality.
NR Blur measurement for VQA.
Lu proposed a method for blur evaluation to mea-
sure blur caused by video compression and imaging processes [64] . First a low pass
filter is applied to each frame in order to eliminate blocking artifacts. Only a subset
of pixels in a frame are selected on the basis of edge intensity and connectivity for
blur measurement, in order to process only that 'type' of blur we are interested in
(for example, blur due to compression as against blur due to a low depth of field).
Blur is then estimated using a combination of an edge image and gradients at the
sample points. The authors demonstrated that their algorithm correlates well with
PSNR and the standard deviation of the blurring kernel for three videos.
NR Fluidity Measure.
In this chapter, we have so far discussed NR measures which
generally do not model frame-drops. The measure proposed by Pastrana-Vidal and
Gicquel in [65] covers this important aspect of NR VQA. Other works along the
same lines that we haven't discussed here include [66], [67] and [68]. In [65] the
discontinuity along the temporal axis, which the authors label fluidity break is first
isolated and its duration is computed. Abrupt temporal variation is estimated using
normalized MSE between luminance components in neighboring frames. Based on
