Video Quality Analysis for Concert Video Mashup Generation - Advanced Concepts for Intelligent Vision Systems - page 4

the luminance component, derived from the YCbCr color-space, of the frame
as it contains most of the blockiness information. For both the horizontal and
the vertical directions, we use the Sobel operator to obtain the gradient image.
Then we apply two thresholds, T h and T l , to the boundary pixels such that if
the gradient value is above T h , the pixel is considered as a real or hard edge,
while if the value is between T h and T l the pixel is considered as a soft edge,
which causes the blockiness effect. The values of T l and T h are chosen as 50
and 150, respectively based on their performance on the test data-set. A block
boundary β ij is considered blocky if more than 75% (6 out of 8) boundary pixels
correspond to soft edges. Blockiness measurement in the horizontal direction B h
is specified by:

H/ 8 − 1 W/ 8 − 1

1

B h = β ij , (1)

( W/ 8 − 1)( H/ 8 − 1)

i =1 j =1

where β ij = 1 if a block boundary is a soft edge and zero otherwise, W and
H represent the number of rows and columns of the frame, respectively. The
vertical blockiness B v is computed in a similar way. The blockiness of a frame
B is computed as the average of both horizontal and vertical blockiness as B =
( B h + B v ) / 2. Fig. 1 shows three frames with different amounts of blockiness.

(a) B =0 . 01 (b) B =0 . 09 (c) B =0 . 17

Fig. 1. Examples of test frames with (a) low, (b) medium and (c) high blockiness. The
measured blockiness is given by B .

2.2 Blurriness

Blurriness is characterized by the reduction of sharpness of edges. Blur can be
caused by lens out of focus or shakiness during capturing, coarse quantization
during compression and filtering for blockiness or noise removal during decoding.
We test methods for blurriness measurements based on features: signal ac-
tivity [4] and average spread of edges [8]. We select the latter method since the
blurriness measurement matches better according to our pilot test results. In this
method, first the edges are detected using the Sobel operator and the number of
pixels contained within the non-zero part of the gradient waveform are counted.
This number is specified as the spread of an edge. More specifically, if there are
m pixels representing an edge and s pixels representing the total edge spread,
then the blurriness score ( Z )iscalculatedas: