Video Quality Analysis for Concert Video Mashup Generation - Advanced Concepts for Intelligent Vision Systems - page 7

Fig. 4 shows a pan speed before and after the low-pass filtering represented by
a thin line and a bold line, respectively.

15

10

5

0

−5

−10

−15

1800 1900 2000 2100 2200 2300 2400

Frame number

Fig. 4. Camera panning speed in a video before filtering ( p ), represented by a thin line
and after filtering ( p f ), represented by a bold line

2.5 Overall Image Quality

The measured values of the quality factors: blockiness, blurriness, brightness,
and shakiness determine the quality of a video frame. Since the values corre-
sponding to the different factors have different ranges, we normalize them with
respect to their corresponding maximum values obtained from test recordings.
After normalization, the values of the factors are in the range [0,1]. The values
obtained from the method are optimized for the test-set. In ideal case, the values
is chosen based on the experimental results with thousands of concert video in
different conditions, such as indoor, outdoor, various camera types. However,
due to practical limitations we focused on the videos obtained from YouTube.
There are different possible methods to combine the factors to estimate the
image quality, such as linear addition [3], linear multiplication [4] and non-linear
fusion [5]. No consensus is found in the prior works [3]-[5] over the weights of
different quality factors on the overall quality measurement of an image. Based
on our observation of the concert recordings and the pilot test, we consider
that all the quality factors are equally important. We test the linear addition
and multiplication methods on overall quality score. Since the quality score us-
ing multiplication method has a wider range than addition method, it is more
suitable for comparing the scores of different recordings in a multiple-camera
recording. Therefore to compute the overall quality score of a frame, we use the
product of the different quality scores. Since shakiness, blurriness and blocki-
ness attribute a negative quality, the factor values are subtracted from one. The
image quality score q of a video frame is given by:

J ) , (8)
where I ,B ,Z and J represent the normalized values of the brightness, block-
iness, blur and shakiness, respectively.

q = I (1

B )(1 Z )(1 − − −