Game Development Reference
In-Depth Information
To this end, in the literature various global features are extracted to represent the
image quality. The accuracy of the quality prediction is dependent on the extracted
features. In Wang and Simoncelli (
2005
), the distance between the probability dis-
tributions of wavelet coefficients is calculated to measure the perceptual quality. The
differences between the entropies of wavelet coefficients of reference and distorted
images are quantified to measure the image quality (Soundararajan and Bovik
2012
).
Based on the natural scene statistics, in Wang et al. (
2006
) the authors predict the
image quality by measuring the destruction of “naturalness.” In Ma et al. (
2011
), the
DCT coefficients are reorganized into several representative sub-bands and the image
quality is evaluated based on the city-block distance. Apart from these approaches,
predicting the full reference metric from extracted statistical features is also recog-
nized as an effective scheme. For example, in Rehman and Wang (
2012
) the SSIM
index is predicted from a multiscale multiorientation divisive normalization trans-
form.InGuetal.(
2013
), the distance between the structural degradation information
of the original and distorted images is used to quantify the image quality. In Wang
et al. (
2013
), the features derived from the entropy of primitives (EoP) are employed
to evaluate the image quality.
Recently, the free-energy principle is proposed by Friston et al. (
2006
), Friston
(
2010
), which makes a basic assumption that the cognitive process is controlled by
an internal generative model in the human brain. This has attracted many attentions
in RR IQA. The underlying idea of the free-energy principle is that all adaptive
biological agents resist a natural tendency to disorder, and attempts to maintain their
internal states at the low entropy level so as to avoid surprises. Therefore, it makes
sense to deal with the primary information and uncertainly separately. In Wu et al.
(
2013
), the primary visual information and the residual uncertainty are combined to
evaluate the image quality. In Zhai et al. (
2012
), the linear AR model is chosen as the
generative model for its ability to approximate a wide range of natural scenes, and
the image quality is defined to be the discrepancy measure between the input image
and its best explanation inferred by the internal generative model.
11.1.4 No Reference Image Quality Assessment
Compared with the RR method, blind quality assessment or NR IQA is usually more
difficult as no information about the reference image can be relied on. Therefore,
a realistic objective is to conduct NR image IQA focusing on artifacts introduced
by a certain kind of image processing algorithm. For example, in Ong et al. (
2003
)
the image quality is quantified by the average edge spread. The slope's spreads of
the extracted edges are measured, and then a linear transform is used to infer the
quality of JPEG2000 images. In Li (
2002
), both the sharpness and ringing artifacts
are measured in edge sharpness level, noise level and random noise level. The edge
sharpness is characterized by the estimations of the scale parameters of an ideal two-
dimensional step edge and the blur artifacts are evaluated through the scale parameter
histograms. The ringing artifacts are measured in the structural noise level. With the
