Information Technology Reference
In-Depth Information
a
1
DMOS
=
(3)
P
1
+−⋅
exp(
aIQa
(
−
))
2
3
The nonlinear regression function is used to transform the set of metrics results to a
set of predicted DMOS values,
DMOS
P
, which are then compared against the actual
subjective scores (DMOS) and resulted in three evaluation criteria: root mean
square error (RMSE), Pearson correlation coefficients, and Spearman rank order
correlation coefficient. RMSE and Pearson correlation express the prediction accu-
racy of a quality metric, and Spearman rank order correlation provides information
about the prediction monotonicity of the metric [30]. In addition, the reference [30]
also suggests another criterion, the outlier ratio, that relates to prediction consis-
tency based on standard errors of the subjective quality values. However, the LIVE
dataset does not provide such standard errors for computing the outlier ratio.
Table 2
Performance evaluation of PSNR and SSIM in digital cinema setup and LIVE
dataset
Digital Cinema
LIVE Dataset
Criteria
PSNR
SSIM
PSNR
SSIM
RMSE
1.00
1.13
7.45
5.71
Pearson
0.914
0.888
0.888
0.936
Spearman
0.913
0.875
0.890
0.931
Table 2 gives the performance evaluation of PSNR and SSIM on image quality as-
sessment in our digital cinema setup and LIVE image dataset, respectively. It is
noticed that the RMSE values are strongly related to score ranges in a subjective
quality assessment, which is why the RMSE values between the digital cinema
scenario and LIVE dataset are quite different. However, according to the compari-
son on the correlation coefficients, we can find that the performance of SSIM is
worse than PSNR. In our opinion, the reason is that subjects do not compare the
entire structural information between the distorted image and the reference image
in a digital cinema setup, because the image size is too large.
According to the above performance comparison of PSNR and SSIM between
two different scenarios, as well as the analysis of the characteristics of the HVS in
a digital cinema setup, we think that an image quality metric developed for normal
size images cannot be adopted as is to predict the image quality in digital cinema
applications. The main reason is that a subject cannot see an entire image at once
when he/she evaluates the quality of this image in digital cinema. Thus, we pro-
pose to first divide the image into different blocks, and then perform image quality
metrics on each block. The overall quality of this image can be derived from the
metric results in all blocks or parts of the image blocks. In this study, square
blocks are employed. The block size (
S
) is determined based on several factors,
including the human fovea acuity angle (
α ), image size (
S
1
), screen size (
S
2
), and
viewing distance (
D
), as follows:
