Information Technology Reference
In-Depth Information
S
α
1
S
=⋅ ⋅
D
atan(
)
(4)
S
180
2
In the subjective quality assessment introduced in Section 4, the reference and dis-
torted images were shown on the screen simultaneously (Fig. 3), therefore, we
used the height values of the image and the screen as their sizes
S
1
and
S
2
.
The reference and the distorted images have been divided into different blocks
whose size are
S
×
S
, and then PSNR and SSIM values were calculated in each
block between the reference and the distorted images. Actually, it is unnecessary
to perform the computations for all blocks. One reason is that some blocks with
severe distortions may dominate the overall quality of an image [28]. Another
reason is that subjects usually estimate their judgment of the quality of an image
based on evaluation of a subset of regions or blocks in that image, and they may
not have enough time to examine all blocks or regions in a short viewing duration,
such as 10 seconds in our subjective experiments. Furthermore, we found that sub-
jects paid more attention to those image blocks with higher contrast when assess-
ing the image quality. Hence, the standard deviation in each divided block was
computed to express the contrast information, and all the blocks were sorted in a
descending order according to their standard deviation values. Blocks with lower
contrast levels were excluded from the quality calculation, where a threshold
T
for
distinguishing the blocks was set. This threshold (
T
) was estimated, as in equation
(5), according to the saccadic and fixation time of eye movement and the viewing
duration in the subjective quality assessment:
10(s) 30 (ms)
T
=
(5)
M
where
M
denotes the number of all divided blocks in an image, 10(seconds) is the
viewing duration, and 30(milliseconds) is the saccadic and fixation time.
PSNR and SSIM measures are computed in the candidate blocks between the
reference and distorted images whose contrast levels exceed the threshold, and
then mean values of PSNR and SSIM over these blocks were used as the quality
of that image measured by PSNR and SSIM, respectively. Our experimental re-
sults demonstrate that the performance of this approach is better than the original
PSNR and SSIM methods.
As aforementioned, we computed the quality values in those blocks with high
contrast levels. However, the distortions introduced to image blocks, especially
the blocks with high contrast levels, are not perceived by the HVS totally, because
of the contrast sensitivity of the HVS that varies with different frequencies and the
existence of masking effects. Contrast sensitivity is a measure of the ability to dis-
cern between luminance of different levels in an image. In addition, when an im-
age has high activity, there is a loss of sensitivity to errors in those regions. This is
the masking effect of the image activity. Many approaches have been proposed to
model the contrast sensitivity and masking effects in order to compute a visually
optimal quantization matrix for a given image in compression algorithms. The
Discrete Cosine Transform (DCT) has usually been used in contrast making due to
its suitability for certain applications and accuracy in modeling the cortical
