Information Technology Reference
In-Depth Information
Fig. 8 (Cont.)
(e) Distorted image with white noise and disparity
Table 6 Evaluation results of image quality metrics on disparity images
Criteria
GCC
MSE
MAD
PSNR
SSIM
MSSIM
VSNR
RMSE
7.11
7.09
7.22
6.86
6.40
6.99
6.85
Pearson
0.826
0.827
0.820
0.839
0.862
0.832
0.840
Criteria
VIF
UQI
IFC
NQM
WSNR
PHVS
JND
RMSE
8.31
6.37
8.08
7.94
7.20
6.87
7.59
Pearson
0.755
0.863
0.769
0.776
0.821
0.839
0.816
As explained above, the disparity refers to the difference in location of an object
seen by the left and right eyes. Thus, the disparity images have quite different mo-
dalities compared to the original images, as shown in Fig. 8. Firstly, we tested three
simple metrics: global correlation coefficient (GCC), mean square error (MSE),
and mean absolute difference (MAD). We performed the same fitting operation, as
in Equation (2), between the computed results obtained by these metrics and the
DMOS values on the whole distortion types, and then the Pearson correlation coef-
ficient and RMSE were calculated. Secondly, we also validated the performance of
IQMs on the disparity images, even though these IQMs were supposed to be devel-
oped for predicting the quality of natural images. Table 6 gives the evaluation
results of the Pearson correlation coefficient and RMSE using these metrics.
According to the evaluation results of the IQMs on the disparity images, the
performance is much better than that on original images. This observation proba-
bly indicates that the disparity information is more important than the original im-
ages for perceptual quality assessment, even though the disparity does not contain
any real objects. The big differences between two disparity images usually appear
in the regions where the distortions are greatly annoying. Thus, even a very simple
metric on the disparity images, such as MSE, performs better than a complicated
IQM on the original images. Additionally, we found that SSIM and UQI have the
best performance within all the IQMs. We believe that this is because these two
 
Search WWH ::




Custom Search