Information Technology Reference
In-Depth Information
metrics are based on comparing the structural information, and the disparity can
express such structural information of the original images.
Since the disparity images have significant influence on the stereoscopic image
quality assessment, we naturally suppose that the combination of the disparity im-
ages and the original images can perform better than using either the disparity or
the original images solely. Subsequently, we used three approaches to combine the
disparity and original images to compute the stereoscopic image quality.
Table 7
Evaluation results of global combination between image quality and disparity
quality on stereoscopic image quality assessment
IQ
DQ
PSNR SSIM MSSIM VSNR VIF UQI IFC NQM WSNR PHVS JND
GCC 0.869 0.867 0.840 0.830 0.831 0.835 0.836 0.837 0.828 0.833 0.839
MSE 0.887 0.878 0.838 0.830 0.828 0.844 0.843 0.829 0.847 0.828 0.846
MAD 0.888
0.899
0.853 0.828 0.825 0.841 0.833 0.829 0.838 0.830 0.851
PSNR 0.876 0.887 0.848 0.836 0.837 0.847 0.874 0.842 0.840 0.839 0.829
SSIM 0.858 0.859 0.870 0.862 0.858 0.870 0.861 0.866 0.856 0.859 0.866
MSSIM 0.857 0.865 0.837 0.832 0.836 0.846 0.853 0.833 0.840 0.834 0.815
VSNR 0.850 0.842 0.844 0.841 0.837 0.860 0.834 0.838 0.833 0.845 0.863
VIF 0.817 0.819 0.804 0.741 0.779 0.826 0.730 0.732 0.730 0.766 0.778
UQI 0.855 0.859 0.865 0.862 0.858 0.868 0.863 0.868 0.855 0.857 0.864
IFC 0.814 0.807 0.793 0.764 0.775 0.822 0.760 0.762 0.760 0.778 0.780
NQM 0.847 0.856 0.829 0.770 0.784 0.827 0.774 0.764 0.763 0.796 0.775
WSNR 0.865 0.878 0.852 0.817 0.831 0.838 0.840 0.821 0.818 0.823 0.818
PHVS 0.853 0.879 0.845 0.813 0.818 0.843 0.823 0.817 0.813 0.818 0.825
JND 0.839 0.876 0.827 0.833 0.806 0.852 0.795 0.836 0.827 0.815 0.839
The first approach, called global combination, was to compute two quality val-
ues of the distorted image and the distorted disparity firstly, denoted as
IQ
and
DQ
, respectively.
IQ
was computed by IQMs on the original images, and
DQ
by
GCC, MSE, MAD, and the IQMs. Then, an overall quality which was taken as the
quality of the stereoscopic image was calculated using the following function with
different coefficients and exponents:
OQ
=⋅ +⋅ +⋅ ⋅
a IQ
d
b DQ
e
c IQ
d
DQ
e
(3)
In this study, we employed Levenberg-Marquardt algorithm to find the optimum
parameters in Equation (3). Although the optimum parameters may change if dif-
ferent initial values were used, we found that the highest correlation coefficient
between
OQ
and DMOS values is 0.899. For example, one set of the optimum pa-
rameters is
a
=3.465,
b
=0.002,
c
=-0.0002,
d
=-1.083, and
e
=2.2. In this experiment,
we used the direct correlation between
OQ
and DMOS values while the fitting op-
eration in Equation (2) was not performed because we have performed an
