Information Technology Reference
In-Depth Information
optimization operation between
OQ
and DMOS values in Equation (3). We report
the highest correlation for different combinations in Table 7 while the correspond-
ing optimum parameters are omitted for the sake of clarity.
According to the experimental results, it was found that appropriate combina-
tions of the image quality and the disparity quality perform better than using the
quality of either the original images or the disparity images solely. In addition, we
also found that the combination of SSIM and MAD, i.e. SSIM was used to compute
IQ
and MAD was used to compute
DQ
, always obtains the best performance within
all the possible combinations. Furthermore, SSIM has a promising performance
in the combinations either for measuring the original image quality or for comput-
ing the disparity image quality. This result indicates that a good metric for predict-
ing the stereoscopic image quality can be developed if appropriate methods are
found to combine the original image quality and the disparity image quality.
The second approach is called local combination. Some IQMs, e.g. PSNR
(based on MSE), SSIM, MSSIM, UQI, PHVS, and JND, compute a quality map
between the reference image and the distorted image to depict the distribution of
quality degradation at image pixels directly or indirectly, and the overall quality of
the distorted image is usually computed as a mean over all the pixels in the quality
map. Furthermore, we can also compute a quality map of the disparity image
which can reflect an approximate distribution of the degradation on the distorted
disparity image. In this study, four methods were used to compute the quality map
on the disparity image as following:
⎧
(
DD
−
)
2
⎪
⎪
DD
−
⎪
DDQ
=
(4)
⎨
2
DD
2
−
⎪
1
−
⎪
⎪
⎩
255
(, )
IMQDD
where
D
and
D
denote the orig
in
al disparity image and the distorted disparity
image, respectively, and
IMQDD
denote the quality map using the correspond-
ing IQMs (including PSNR, SSIM, MSSIM, UQI, PHVS, and JND) between the
original disparity image and the distorted disparity image. After computing the
quality maps of the original image and the disparity image, Equation (3) was used
to pool each pixel pair on the quality maps, and then the mean over all pixels was
taken as the overall quality of the stereoscopic image. Table 8 gives the Pearson
correlation coefficients between the quality values and the DMOS values by using
the local combination, where the highest correlation coefficients were reported.
According to the evaluation results, it was found that the performance im-
provement by using the local combination is not as significant as if the global
combination was performed. Some combinations even reduced the correlation be-
tween the overall quality values and the subjective DMOS values. However, we
found that SSIM and UQI algorithms on the original images and disparity images
have the best performance for local combination, regardless of what kinds of
(, )
