Game Development Reference
In-Depth Information
combination of measurements in all the levels, the overall image quality is appraised.
In Marziliano et al. (
2004
), both the full-reference and NR blur metrics are modeled
by the spread of the vertical edges, and the full-reference ringing metric is also
modeled by the characteristics of the edge-spreads. In Barland and Saadane (
2005
),
the blur artifact measurement depends on the spatial information on the whole image,
while the ringing artifact measurement only depends on the local information around
strong edges.
Another branch is based on the assumption that the compression process may
disturb the naturalnesswithin the natural scenes, and the disturbance can be quantified
to predict the human perceptions of image quality. Some representative approaches
are as follows. In Sheikh et al. (
2005
), the fact that the quantization in JPEG2000
coding pushes wavelet coefficients at finer scales toward zero is considered. A natural
scene statistics (NSS) model is built to model the wavelet coefficient's magnitude
conditioned on the magnitude of the linear prediction of the coefficient. The sub-
band probabilities computed based on the model are used to indicate the loss of visual
quality. In Shaked and Tastl (
2005
), a hybrid method based on several previous works
is developed in an efficient form, in which the fractal image model (Pesquet-Popescu
and Véhel
2002
) is involved.
Recently, the general-purpose NR IQA algorithms have also been intensively
studied in recent years. Effective features are firstly extracted from the distorted
images, and then a regression process is adopted to predict the quality. For example,
the natural scene statistics based DIIVINE, BLIINDS-II, and BRISQUE (Moorthy
and Bovik
2011
; Saad et al.
2012
and Mittal et al.
2012
) are conducted IQA in DWT,
DCT, and spatial domain, respectively. In Xue et al. (
2013
), the quality-aware clus-
tering (QAC) is proposed by learning a set of quality-aware centroids as a codebook
to predict the quality score of the overall image. Though various approaches are pro-
posed to infer the visual quality of the distorted image without reference, it is noted
that the NR quality assessment is still in its early stage, and separating the distortion
from content is still a very challenging task.
11.2 Video Quality Assessment
11.2.1 Video Quality Assessment Database
Recently, various video quality assessment databases have been developed for bench-
marking the VQA performance. Popular databases including the EPFL-PoliMI
Video Quality Assessment Database (De Simone et al.
2009
), LIVE Video Quality
Database (Seshadrinathan et al.
2010
), LIVE Mobile Video Quality Database
(Moorthy et al.
2013
), Poly-NYU Video Quality Databases (Ou et al.
2011
), TUM
1080p25 Data Set, TUM 1080p50 Data Set, VQEG FR-TV Phase I Database and
VQEG HDTV Database.
VQEGFR-TVPhase I Database is perhaps the earliest video database that contains
MPEG-2 compression and transmission distortions. It is extended in VQEG HDTV
