Information Technology Reference
In-Depth Information
luminance comparison is computed only at the highest scale. The overall MSSIM
measure is obtained by combining the measures at different scales.
VSNR (Visual Signal-to-Noise Ratio) [33] operates via a two-stage approach.
In the first stage, contrast thresholds for detection of distortions in the presence of
natural images are computed by wavelet-based models of visual masking and vis-
ual summation. The second stage is applied if the distortions are suprathreshold,
which operates based on low-level visual property of perceived contrast and mid-
level visual property of global precedence. These two properties are measured by
the Euclidean distance in a distortion-contrast space of multi-scale wavelet de-
composition. VSNR is computed based on a simple linear sum of these distances.
VIF (Visual Information Fidelity) [34] is to quantify loss of image information
to the distortion process based on natural scene statistics, the human visual system,
and an image distortion model in an information-theoretic framework.
UQI (Universal Quality Index) [35] is similar to SSIM, and it is to model image
distortions as a combination of three factors: loss of correlation, luminance distor-
tion, and contrast distortion.
IFC (Information Fidelity Criterion) [36] is a previous work of VIF. IFC is to
model the natural scene statistics of the reference and distorted images in wavelet
domain using steerable pyramid decomposition [37].
NQM (Noise Quality Measure) [38] is a measure aiming at the quality assess-
ment of additive noise by taking into account variation in contrast sensitivity,
variation in local luminance, contrast interaction between spatial frequencies, and
contrast masking effects.
WSNR (Weighted Signal-to-Noise Ratio) [38] is to compute a weighted signal-
to-noise ratio in frequency domain. The difference between the reference image
and distorted image is transformed into the frequency domain using a 2D Fourier
transform and then weighted by the contrast sensitivity function.
PHVS (PSNR based on the Human Visual System) [39] is a modification of PSNR
based on a model of visual between-coefficient contrast masking of discrete cosine
transform (DCT) basis functions. This model can calculate the maximal distortion that
is not visible at each DCT coefficient due to the between-coefficient contrast masking.
JND (Just Noticeable Distortion) [40] model integrates spatial masking factors
into a nonlinear additivity model for masking effects to estimate the just notice-
able distortion. A JND estimator applies to all color components and accounts for
a compound impact of luminance masking, texture masking and temporal mask-
ing. Finally, a modified PSNR is computed by excluding the imperceptible distor-
tions from the computation of the traditional PSNR.
Because four typical distortion types were adopted in the subjective quality as-
sessment on the stereoscopic images in Section 2, we will also investigate the per-
formance of these IQMs on 2D images with the same distortion types. The source
2D images and corresponding subjective evaluation results were collected from
the LIVE image quality database [15, 31, 41], and the distortions are as following:
•
Gaussian blur: The R, G, and B color components were filtered using a circu-
lar-symmetric 2-D Gaussian kernel of standard deviation σB pixels. These three
color components of the image were blurred using the same kernel. The values
of σB ranged from 0.42 to 15 pixels.
