Information Technology Reference
In-Depth Information
combination are used. For example, let
IQM
and
DQM
be quality maps of the
original image and disparity image computed by UQI and SSIM, respectively, and
the combination at each pixel pair be
OQM
= + + ⋅ ,
the Pearson correlation coefficient between the predictive qualities and the subjec-
tive results is 0.899. Therefore, not all metrics are suitable for the local combina-
tion, and an appropriate method is needed to explore the relationship between the
image quality map and the disparity quality map.
IQM
DQM
IQM
DQM
Table 8
Evaluation results of local combination between image quality map and disparity
quality map on stereoscopic image quality assessment
IQ
DQ
PSNR SSIM MSSIM UQI PHVS JND
2
0.866 0.832 0.776
0.833 0.837 0.770
(
DD
−
)
DD
−
0.840 0.849 0.743
0.853 0.806 0.803
2
DD
2
−
0.807 0.792 0.815
0.898 0.815 0.795
1
−
255
PSNR
0.799 0.805 0.786
0.842 0.776 0.782
SSIM
0.799 0.821 0.800
0.899
0.832 0.816
MSSIM
0.762 0.801 0.774
0.859 0.769 0.802
UQI
0.798 0.822 0.831
0.895 0.841 0.839
PHVS
0.795 0.826 0.804
0.846 0.765 0.728
JND
0.804 0.823 0.781
0.835 0.774 0.807
Finally, the third approach was to integrate the local combination into the
global combination by the following three steps:
•
Two quality maps were computed firstly using appropriate metrics on the origi-
nal image and disparity image, respectively;
•
These two maps were combined locally and the mean was taken as an interme-
diate quality of the distorted image;
•
The final step was to combine the intermediate quality and the quality of the dis-
parity image and then obtained the overall quality of the stereoscopic images.
In our experiment, the highest correlation coefficient (
0.91
) was achieved when UQI
was used in computing the quality maps of the original image and disparity image, and
the local combination on the quality maps was then combined with the MAD of the
disparity image again. Figure 9 gives the scatter plot of the subjective DMOS values
versus the optimum predictive quality values. According to the experiment results, the
proposed model has better performance on predicting perceptual quality of the stereo-
scopic images with lower impairments than that on the images with higher
impairments. Therefore, improving the robustness of the quality metric to different
impairment levels is also an important task in the future work.
