Information Technology Reference
In-Depth Information
focused on certain regions known as salient regions or attention regions [26]. We
have proposed a visual attention based perceptual quality metric by modeling the
visual attention based on a Saliency model and visual content analysis [27]. How-
ever, it was found that the visual attention does not have an evident influence on
the image quality assessment. The NTIA video quality model (VQM) assumes that
the HVS is more sensitive to those regions with severe distortions [28We firstly
tested the performance of two image quality metrics: PSNR and SSIM, on image
quality assessments for normal size images and high-resolution images in a digital
cinema setup. PSNR is a traditionally used metric based on Mean Square Error
(MSE), computed by averaging the squared intensity differences between the dis-
torted and reference image pixels, and defined as follows:
2
P
PSNR
=⋅
10 log
(
)
(1)
10
MSE
where
P
denotes the peak value of the image. Although PSNR does not always
correlate well with subjective image quality assessment, it is still widely used for
evaluation of the performance of compression and transmission systems. PSNR
and MSE are appealing because they are simple to compute, have clear physical
meanings, and are mathematically convenient in the context of optimization.
Based on the hypothesis that the HVS is highly adapted for extraction of struc-
tural information, Wang
et al.
have developed a measure of structural similarity to
estimate image quality by comparing local patterns of pixel intensities that have
been normalized for luminance and contrast. The SSIM measure is calculated
based on three components: luminance, contrast, and structure comparison,
defined as follows:
(2
μμ
+
C
)(2
σ
+
C
)
xy
1
y
2
SSIM x y
(, )
=
(2)
(
μμ
2
++
2
C
)(
σσ
2
++
2
C
)
x
y
1
x
y
2
where
denote mean and standard deviation on the luminance compo-
nent, and
C
1
and
C
2
are two small constants to avoid instability when
and
σ
μ
2
2
(
μμ
+
)
or
x
y
2
2
σ+ are very close to zero. In addition, the authors of SSIM calculated the
SSIM measure within a local square window, moving pixel-by-pixel over the en-
tire image. A mean of SSIM indices over all windows is computed after applying
a circular-symmetric Gaussian weighting function to the reference and the dis-
torted images to eliminate blocking artifacts.
In this study, we used our subjective quality assessment in Section 4 to evaluate
the performance of PSNR and SSIM for digital cinema applications. In addition,
the subjective quality results on JPEG 2000 compressed images with normal sizes
were extracted from the LIVE image quality dataset [29], where 29 reference im-
ages have been compressed using JPEG 2000 at different bit rates ranging from
0.028 bits per pixel (bpp) to 3.15 bpp. After calculating the quality values using
PSNR and SSIM on these distorted images, a nonlinear regression operation be-
tween the metric results (
IQ
) and the subjective scores (DMOS), as suggested in
[30], was performed using the following logistic function:
(
)
x
y
