Information Technology Reference
In-Depth Information
all processing is undertaken using luminance information alone. Color-distortions
and their assessment are interesting areas of research [22].
4
Reduced-Reference Algorithms
Reduced reference (RR) VQA algorithms generally follow a model in which some
amount of information from the source video is embedded in the video being trans-
mitted. This embedded information is recovered at the receiver and compared with
the same information extracted from the received video to assess quality. We clas-
sify these algorithms into two categories - one that relies on inserting a watermark
in each frame and all other techniques including natural scene statistics, multivari-
ate data analysis and so on. Even though some authors claim that the techniques
based on watermarking are NR, based on our definition, embedding information at
the sender in the video stream constitutes a RR scheme.
4.1
Techniques Based on Watermarking
Hidden markers for MPEG quality.
Sugimoto
et. al.
proposed a technique for
measuring MPEG quality based on embedding binary data in videos [23]. A partic-
ular previously agreed upon binary sequence (for example, a series of 1's or 0's or
alternating 1's and 0's) is first generated. The AC components of each macroblock
2
in each frame are computed and its spectrum is spread using a pseudo-random se-
quence (PN) [24]. Embedding is then performed in the frequency domain using a
set of rules based on the quantization parameter. At the receiver a reverse process is
performed and the false detection rate is computed. The authors claim that this rate
is linearly correlated with PSNR and hence with quality.
Tracing watermark and blind quality.
A different PN matrix for each frame is
multiplied with the same watermark, and this product is embedded in the mid-band
frequencies from the discrete cosine transform (DCT) of each frame [25]. At the
receiver a reverse procedure is carried out and the average mean squared error be-
tween the transmitted and the original watermark across the video is computed. Even
though correlation with perceived quality is not evaluated, the authors demonstrate
that mean squared error (MSE) correlates with the bit-error rate.
Fu-Zheng
et. al.
propose a similar method for VQA, however instead of using
the whole frame, they use only certain sub-blocks is alternate frames [26]. Further,
the watermark is weighted based on the quantization parameter and instead of using
MSE, the number of correctly recovered watermark pixels are used as a measure of
degradation. Correlation with PSNR is demonstrated.
2
A macroblock is the default unit of coding for video compression algorithms. Refer to [21]
for details.
