Game Development Reference
In-Depth Information
11.1.1 Image Quality Assessment Databases
In the last decades, there are many image quality assessment (IQA) databases
developed for benchmarking the IQA algorithms. Digital images with various kinds
of distortions and their corresponding subjective ratings are included, which have
greatly facilitated the research of objective IQA metrics in the recent years. Popular
databases including the Laboratory for Image andVideo Engineering (LIVE) (Sheikh
et al.
2005
), Tampere Image Database 2008 (Ponomarenko et al.
2009
), Categorical
Subjective Image Quality (CSIQ) database (Larson and Chandler
2010
), contrast-
changed image quality database (CID2013) (Gu et al.
2013
), and high dynamic range
image quality database (HDR2014) (Liu et al.
2014
), Tampere Image Database 2013
(Ponomarenko et al.
2013
), image retargeting database (Ma et al.
2012
), and LIVE
multiply distorted image database (LIVEMD) (Jayaraman et al.
2012
).
LIVE database covers five distortion types including white noise, JPEG2000,
JPEG, Gaussian blur and fast fading. In total, 779 distorted images are created from
29 pristine ones with these distortion types at different levels. The TID2008 database
was developed as a joint effort between countries such as Finland, Italy, and Ukraine.
It includes 1,700 distorted images generated from 25 reference images with 17 dis-
tortion categories. For each distortion type, four distortion levels are generated for
each reference image. The categorical image quality (CSIQ) database was developed
at Oklahoma State University and consists of 866 images which are derived from
30 reference images. Six distortion types with four to five levels are contained in
CSIQ. For contrast change images, a more comprehensive database was developed
(CID 2013), which consists of 15 natural images taken fromKodak database and 400
contrast-changed versions. Luminance shifts and contrast changes are used to gener-
ate the distorted images. Recently, high dynamic range (HDR) imaging has attracted
a lot of attentions in both academic and industry. A new and dedicated HDR image
quality database (HDR 2014) was proposed, which contains four kinds of distortion
types and eight distortion levels. The TID2013 database was as an extension of the
TID2008 database, which contains 3,000 images generated by corrupting 25 original
images with 24 types of distortion at five different levels. Recently, content-aware
image retargeting has attracted a lot of interests. The image retargeting quality data-
base contains 171 retargeted images (obtained from 57 natural source images of
different contents) by several representative image retargeting methods. The above-
mentioned databases are all single type distortions, which may not cover the real
world distortions as different distortions tend to occur together in one image. This
motivated the LIVE Laboratory to create a multiply distorted image database. LIVE
multiply distorted database was created by two distortion scenarios: (1) image stor-
age: where images are first blurred and then compressed by a JPEG encoder; (2)
camera image acquisition: where images are first blurred due to defocus and then
corrupted by white Gaussian noise.
