Content uploaded by Vipin Kamble
Author content
All content in this area was uploaded by Vipin Kamble on May 02, 2016
Content may be subject to copyright.
No-Reference Image Quality Assessment Algorithms:
A Survey
Vipin Kamble
Email id - vipinkamble97@gmail.com
Department of Electronics Engineering,
Visvesvaraya National Institute of Technology, Nagpur, India
K. M. Bhurchandi
Department of Electronics Engineering,
Visvesvaraya National Institute of Technology, Nagpur, India
Abstract:
Evaluation of noise content or distortions present in an image is same as assessing the quality of
an image. Measurement of such quality index is challenging in the absence of reference image.
In this paper, a survey of existing algorithms for no-reference image quality assessment is
presented. This survey includes type of noise and distortions covered, techniques and parameters
used by these algorithms, databases on which the algorithms are validated and benchmarking of
their performance with each other and also with human visual system.
Keywords: image analysis, image quality, no reference image quality assessment.
1. Introduction
Estimation of noise content of a signal/image and its subsequent removal is a very important area
of research. Till these years, human intelligence was considered to be the only tool for sensing
noise in signal/image. However, a few developments in signal processing like transform based
statistical tools have generated a ray of hope that noise sensing by machines may become
possible. This paper presents a brief review of major published algorithms for no reference noise
sensing in an image which is referred to as „Image Quality Assessment‟ (IQA). Finally we
present a few parameters for image quality assessment.
Quality of an image represents the amount of visual degradations of all types present in an
image. Degradations may occur due to presence of noise, blocking artifacts, blurring, fading etc.
These degradations are introduced during image acquisition, compression, storage, transmission,
decompression, display or even printing. Sensing the degradation at the time of image
acquisition can be useful to take counter measure to reduce the degradation while storing the
image as a file. In general, sensing quality of an image during various stages of operations may
be helpful for appropriate reconstruction. It saves unnecessary application of denoising
algorithms. It may also suggest appropriate denoising or processing algorithm to retrieve quality
of degraded image. A noise-sensing or IQA algorithm can provide types of distortions present in
the image along with levels of degradation which can be used while denoising as shown in
Figure 1.
Figure 1: Application of NR-IQA - Image Denoising
Overall image quality cannot be evaluated by only a few parameters like brightness, contrast or
sharpness which can be mathematically calculated from image pixels. A sharp image can have
salt and pepper noise present in it thereby devaluating its quality. A standardized evaluation
procedure is required to assess the quality of an image irrespective of the type of distortions
present. This evaluation procedure and the results should confirm well with human perception of
an image quality.
The simplest way to evaluate the quality of an image is to show it to an expert human observer.
However, human perception may be different for each individual. One can tackle this problem by
taking multiple views from different individuals and then statistically processing the results. This
is called subjective image quality assessment. But it is a very lengthy and imprecise procedure
for quality evaluation. Right from selection of observers, their knowledge, expertise, availability,
seriousness bias, interpretations everything is subjective and qualitative. Thus an automated
system for quantitative evaluation of images is required. The problem can thus be represented as
“Quantitative Evaluation of Quality”. Obviously this may require a high level of intelligence.
Automated evaluation of image quality by means of machine is referred to as „Objective Image
Quality Assessment‟. Objective image quality assessment can be accomplished in three ways,
a) Full Reference Image Quality Assessment (FR-IQA)
b) Reduced Reference Image Quality Assessment (RR-IQA)
c) No Reference Image Quality Assessment (NR-IQA)
Full Reference Image Quality Assessment (FR-IQA) refers to assessing the quality of distorted
image by comparing with the original, believed to be undistorted version of same image. The
extent of distortion is calculated by measuring the deviation of distorted image from the
reference image. Simplest way to measure image quality is by calculating the Peak Signal to
Noise Ratio (PSNR) however PSNR does not always correlate with human visual perception and
image quality [1]. To tackle the limitation of PSNR metric, other parameters were proposed.
Parameters which correlate well with human perception include Structural Similarity Index
(SSIM) [2], Visual Information Fidelity (VIF) [3], Fast SSIM [4], Information Fidelity Criteria
(IFC) [5], Multi-scale Structural Similarity (M-SSIM) [6], four-component weighted structural
similarity [7]. These parameters give the extent of deviation of a distorted image from the
reference image. The need of reference image for quality evaluation limits the use of these
parameters and subsequent quality evaluation algorithms.
Reduced Reference Image Quality Assessment (RR-IQA) algorithms are those which use only
limited features from reference image instead of complete image to evaluate the quality of
distorted image. A training approach can also be used for RR-IQA. RR-IQA methods are
mentioned in [8],[9],[10]. The limitation of FR-IQA still remains in RR-IQA i.e., features
extracted from reference image are necessary for quality evaluation. In spite of all its limitations,
the RR-IQA techniques are widely used in satellite and remotely sensed image quality
evaluation.
No Reference Image Quality Assessment (NR-IQA) algorithms provide quality of an image
without the need of any reference image or its features. The problem of NR-IQA is much tougher
than the above two problems. Due to absence of reference image, one needs to model the
statistics of reference image, the nature of human visual system and effect of distortions on
image statistics in an unsupervised way. It is also very difficult to evaluate the effectiveness of a
quality measure with a specific distorted image in absence of a reference image.
This paper provides an extensive review of major NR-IQA algorithms developed so far. This
review summarizes the methods used by algorithms, evaluation parameters, distortion types for
which they are designed, image databases used for validation and benchmarking of the
algorithms with human visual performance.
2. Benchmarking parameters
Different NR-IQA algorithms provide different quality score. So to compare the performance of
different NR-IQA algorithms, a common benchmarking system is necessary. Few benchmarking
parameters are mentioned below:
a) Pearson Correlation Coefficient (PCC)
It is used to measure the dependency between to variables. Its value lies between „-1‟ and
„+1‟ where value close to „+1‟ indicates that the two variables have positive correlation and
value close to „-1‟ indicates that the two variables have negative correlation. A very low or
zero value implies that the two variables are not correlated. Pearson correlation coefficient
(ρ) between two variables „X‟ and „Y‟, with standard deviation σ
x
and σ
y
respectively is
shown in (1).
(1)
The two variables „X‟ and „Y‟ are the output of NR-IQA algorithm and the actual quality
score provided with the database.
b) Spearman Correlation Coefficient (SCC)
Spearman correlation provides the relation between two ranked variables. Its range is from „-
1‟ to „+1‟ with same interpretation as that of Pearson‟s correlation coefficient. Spearman
correlation coefficient is calculated as shown in (2).
(2)
where „d‟ is the difference in ranks of two variables „X‟ and „Y‟.
c) Outlier’s Ratio (OR)
Outlier‟s ratio is defined as the percentage of algorithm‟s output which is beyond twice
standard deviation of subjective scores. If there are „I‟ images and „S
i
‟ is the subjective
quality score of i
th
image then mean subjective score (S
m
) is calculated as shown in (3).
(3)
The standard deviation of subjective scores (σ
s
) is calculated as shown in (4).
(4)
Suppose there are „P‟ images with individual objective quality score (O
i
) such that
(5)
then the outlier‟s ratio is given by (6)
(6)
Outlier‟s ratio increases if the output of NR-IQA algorithm is not at all in agreement with
standard or subjective quality score.
d) Root Mean Square Error (RMSE)
Root Mean Square Error (RMSE) is used to measure the pixel wise deviation between two
entities. If „I‟ and „I
N
‟ are original and noisy images of size MxN, then the RMSE between
these two images is defined as
(7)
3. No-reference Image Quality Assessment Algorithms:
The NR-IQA algorithms that are developed by researchers are discussed in this section. The
algorithms are discussed in the order of their publication.
Pina Marziliano et al. [11] proposed a perceptual blur and ringing metric for images corrupted by
distortion caused by JPEG-2000 compression. The method is based on edge detection using
Sobel operator. Noise and other subtle edges are removed by applying threshold to gradient
image. Edge width per edge is used as a local blur metric. Local blur scores are aggregated to get
final blur present in the image. JPEG-2000 images from LIVE database [12] are used for
validation of the algorithm. The linear correlation coefficient obtained is 73% and Spearman‟s
correlation coefficient is 81%. This metric is designed only for JPEG-2000 compressed images
and tackles only two distortion categories with a low correlation value.
Du Juan et al. [13] presented a quality metric based on model of Human Visual System (HVS).
It uses the concept of jumping phenomenon of eyes which changes point of fixation. The used
HVS model is proposed using fixation point instead of Contrast Sensitivity Function (CSF).
Block size of 10x10 is used for computation. The performance of algorithm is compared using
PSNR metric. Lena image is used for experimentation. This metric is not validated on any
database but only on a few degraded versions of Lena image.
H. R. Sheikh and A. C. Bovik [14] proposed no-reference image quality metric using Natural
Scene Statistics (NSS) of JPEG-2000 compressed images. JPEG-2000 compression disturbs the
nonlinear dependencies which are observed in natural images. Wavelet coefficients‟ magnitudes
and magnitudes of linear prediction of coefficients in four sub-bands are used as statistical
features. Algorithm is trained using Mean Opinion Score (MOS) provided with LIVE database
[12]. Pearson‟s correlation coefficient obtained is 0.91. This metric is designed only for JPEG-
2000 compressed images.
R. V. Babu et al. [15] presented an IQA model for JPEG compressed images. This model uses
Growing and pruning radial basis function (GAP-RBF) [16] for image quality evaluation. The
HVS features used are edge amplitude, edge length, background activity and background
luminance. Background activity refers to amount of high frequency components and background
luminance refers to the amount of brightness. Relation between HVS features and MOS are
modeled using GAP-RBF network. This network uses sequential learning without modifying
past learning so it needs less memory and less computation. In JPEG compression, blocking
artifacts occur at boundaries of 8x8 blocks so same block size is used for patch wise operation.
JPEG compressed images from LIVE database [12] are used for validation. 154 images are used
for training and 79 images are used for testing. The RMSE between Mean Opinion Score (MOS)
and predicted quality score is 0.57. MOS is the average subjective quality score of an image.
This metric is designed only for JPEG compressed images.
Izak van Zyl Marais and Willem Herman Steyn [17] proposed a blur identification metric based
on variation of spectral subtraction method which uses power spectrum surface of revolution.
This method is useful while differentiating between in-focus and out-of-focus blur. Blur is
modeled by a uniform point spread function with a 2D circular support. The algorithm is tested
on 210 images which are obtained by degrading 5 remote sensing images. This metric is used
only for quality evaluation of blurred images and is not validated on any standard database.
Salvador Gabarda and Gabriel Cristóbal [18] used anisotropy as a measure of image quality.
Anisotropy means having different properties in different directions. Generalized Renyi entropy
[19] and the normalized Pseudo-Wigner Distribution (PWD) is used to calculate directional
entropy of an image. Variance of expected entropy is taken as an indicator of anisotropy. Eight
point window is used for Renyi entropy calculation. Images from LIVE database [12] are used
for validation. Standard benchmarking parameters are not used to demonstrate the performance
of this algorithm.
Z.M. Parvez Sazzad et al. [20] proposed a NR-IQA algorithm for JPEG-2000 compressed images
in spatial domain. This method uses pixel distortion and edge information for quality evaluation.
Pixel distortion is measured using standard deviation of central pixel in a 5x5 neighborhood with
partial overlap. Edge information is inferred using zero crossing rate in horizontal and vertical
direction. Pixel distortion and edge information are combined as mentioned in [21]. LIVE
database [12] is used for validation of which, 50% images are used for training and remaining
50% are used for testing. Achieved Pearson‟s correlation coefficient, Spearman‟s correlation
coefficient and outlier ratio are 0.93, 0.99 and 0.0396 respectively. This metric is designed for
quality evaluation of only JPEG-2000 compressed images.
T. Brandao and M. P. Queluz [22] proposed a new method for image quality assessment. The
method uses Natural Scene Statistics (NSS) of Discrete Cosine Transform (DCT) whose
distribution is modeled by Laplacian Probability Density Function [23]. A distribution parameter
is specified for a particular frequency pair. Images from LIVE database [12] is used for testing
and validation. Pearson‟s correlation coefficient and Spearman‟s correlation coefficient achieved
by this technique are 0.973 and 0.978 respectively. This method is trained and tested on only
JPEG coded images from LIVE database. Performance of the algorithm on other image
distortions is not mentioned.
G. Zhai et al. [24] introduced a measure to estimate blockiness in block DCT coded images.
Blockiness refers to pixel distortion caused at block boundaries during encoding of an image.
Blockiness caused by quantization is mapped by block discontinuity. A Noticeable Blockiness
Map (NBM) is obtained from luminance and texture masking. Luminance masking is measured
as luminance difference between neighboring blocks and texture masking is computed using
neighboring blocks directional properties. The algorithm is validated on LIVE [12] and
IRCCyN/IVC [25] databases. Achieved Pearson‟s and Spearman‟s correlation coefficients on
IRCCyN/IVC JPEG database are 0.9688 and 0.9629. Pearson‟s and Spearman‟s correlation
coefficients on LIVE JPEG database are 0.9621 and 0.9087 respectively. As this metric is
designed for blockiness estimation, it can be used only for JPEG and JPEG-2000 compressed
images.
S. Suresh et al. [26] proposed a machine learning approach for NR-IQA of JPEG coded images.
HVS features such as edge amplitude, edge length, background activity and background
luminance are used. These features along with MOS of images are fed to Extreme Machine
Learning (ELM) [27] algorithm to obtain a functional relationship. 154 and 79 images are used
for training and testing respectively. RMSE between Mean Opinion Score (MOS) and proposed
method‟s output is obtained as 0.70. As this is a training based approach, quality assessment of
images distorted with unknown distortion type is unpredictable.
R. Ferzli and L. J. Karam [28] presented a sharpness metric based on concept of Just Noticeable
Blur (JNB). JNB comes from Just Noticeable Difference (JND) [29] and is the minimum
difference in intensity value relative to background that is noticeable. JNB is the minimum
perceivable blur around an edge given a higher contrast than a JND. Standard deviation
corresponding to JNB threshold is calculated using subjective evaluation. Subjective scores are
obtained from the LIVE database [12]. For validation, images corrupted by Gaussian blur and
JPEG-2000 compression are used. Achieved Pearson‟s and Spearman‟s correlation for Gaussian
blur distortion is 0.932 and 0.936 respectively. Pearson‟s and Spearman‟s correlation for JPEG-
2000 compressed images is 0.881 and 0.873 respectively. This method targets only sharpness as
a parameter for image quality evaluation and is validated on blurred and JPEG-2000 compressed
images.
Shiqian Wu et al. [30] proposed a method for assessing the amount of blur present in an image.
The method uses Sobel operator for edge detection and then applies Radon transform to locate
line features. The line spread function and point spread function are calculated from the located
line features. For validation, 13 natural world images are blurred with standard blur levels and
then used for testing the algorithm. Effect of increasing image blur on the output of algorithm is
used to estimate the performance of algorithm. This metric is designed only for quality
assessment of blurred images.
Shan Suthaharan [31] presented a metric for measurement of blocking artifacts caused by block
coding in images. It uses the multi-neural channel aspect of HVS. In this, the primary edges and
undistorted edges are estimated and edges caused by block compression are filtered out. The
filtered edges give an estimate of blocking artifact. JPEG compressed images from LIVE
database [12] are used for validation of algorithm. Pearson‟s correlation coefficient value of 0.91
is obtained. As this metric estimates strength of the blocking artifacts, it is used for quality
evaluation of JPEG and JPEG-2000 compressed images only.
Wen Lu et al. [32] proposed a contourlet transform based Natural Scene Statistics (NSS) model
for image quality assessment. The statistics of contourlet coefficients are described by a joint
distribution function. An image is decomposed in multi-scale and directional sub-bands using
contourlet transform. The algorithm is trained using statistics of images from LIVE database
[12]. Same database is used for validation and the Pearson‟s correlation coefficient is found to be
0.8271. Though the overall performance of algorithms is acceptable but the correlation for JPEG
compressed images is very low, i.e. 0.5810.
Hantao Liu et al. [33] introduced a metric to measure perceived ringing artifact. In this method,
bilateral filtering is used to smooth edges which do not contribute to perceivable ringing. These
edges are obtained by Canny edge detector, skeletonizing, edge linking, noise removal and line
segment labeling. An extracted perceptual edge map obtained from line segments is used to
select edges around which perceivable ringing can occur. Images from “Kodak lossless true color
image suit” are used for validation. Pearson‟s and Spearman‟s correlation coefficients obtained
are 0.868 and 0.85 respectively. This metric can be used for quality evaluation of images which
are distorted by ringing artifact only.
A. K. Moorthy and A. C. Bovik [34] presented a two step framework for image quality
assessment. The algorithm is named as Blind Image Quality Index (BIQI). This method first
estimates the presence of set of distortions and then makes a probability weighted summation to
give a quality score. Wavelet transform [35] over three scales and three orientations is applied to
obtain sub-band coefficients. These coefficients are parameterized using a generalized Gaussian
distribution. Feature vectors obtained from sub-band coefficients are fed to a Support Vector
Machine (SVM) [36] for classification of different distortion types. LIVE database [12] is used
for training and testing of this algorithm. Pearson‟s correlation coefficient, Spearman‟s
correlation coefficient and RMSE are obtained as 0.8205, 0.8195 and 15.6223 respectively. This
algorithm performs well only for distortion types on which the algorithm is trained.
M. A. Saad et al. [37] proposed BLind Image Integrity Notator using DCT Statistics (BLIINDS)
algorithm. It uses Discrete Cosine Transform (DCT) to extract contrast and structural features
from an image. Contrast is obtained from average value of DCT coefficients of 17x17 size
patches. Kurtosis of same size patches is used to obtain DCT based structural features. The
algorithm is trained and tested on LIVE database [12]. Spearman‟s correlation coefficient is
obtained as 0.7996. This is a training based algorithm and is not benchmarked using Pearson‟s
correlations coefficient.
Jing Zhang and Thinh M. Le [38] introduced a quality metric for images coded by JPEG-2000
compression. It uses the monotonically changing pixel activity along horizontal and vertical
directions. The distortion in the monotonicity is considered as degradation. JPEG-2000
compressed images from LIVE database [12] are used for validation of algorithm. Achieved
Pearson‟s correlation and Spearman‟s correlation coefficient are 0.928 and 0.919. This metric is
designed only for JPEG-2000 compressed images and is not generalized.
Luhong Liang et al. [39] presented a measure for assessing the quality of JPEG-2000 compressed
images. This quality metric is designed by combination of a blur and a ringing metric. Blur
metric uses the gradient profile along edges along with the Just Noticeable Difference (JND)
[29]. Ringing metric is characterized by local variance in a small neighborhood region. Weighted
Minkowski summation is used to combine the two metrics i.e., blur metric (M
blur
) and ringing
metric (M
ringing
) as shown in (8).
(8)
The optimum values for Minkowski parameters are 0.85, 0.15 and 3 for „a‟, „b‟ and „p‟
respectively. JPEG-2000 compressed images from LIVE database [12] and TID dataset [40] are
used for validation. Achieved Pearson‟s and Spearman‟s correlations on LIVE database are
0.947 and 0.912 respectively. Pearson‟s and Spearman‟s correlation on TID dataset are 0.911
and 0.902 respectively. This metric is specifically designed for quality evaluation of only JPEG-
2000 compressed images.
Erez Cohen and Yitzhak Yitzhaky [41] proposed a metric to measure blur and noise impact on
images. The image is modeled by Fourier transform, blurring point spread function and additive
noise. Noise distortion is computed in both spatial and frequency domain. In spatial domain,
10*10 size patches are subjected to variance calculation and in frequency domain, image power
spectrum is used. Effect of increasing Gaussian noise variance and defocusing blur diameter on
output of proposed metric is observed for validation. This method is tested on 75 natural
monochrome images [42] but a correlation parameter is not provided for performance
comparison with other algorithms.
Alexandre Ciancio et al. [43] introduced an algorithm for blur assessment using multi-feature
classifier. Outputs of multiple algorithms which include frequency domain metric [44], spatial
domain metric [45], perceptual blur metric [11] and HVS based metric [46] are given to a
classifier. Other features provided to the classifier are Local Phase Coherence using wavelet
transform, mean brightness level, variance of HVS frequency responses [47] and depth of
contrast. The classifier generates an input-output map using a training dataset. Pearson‟s and
Spearman‟s correlation for real blur images is 0.564 and 0.56 respectively. Real blur is difficult
to model so the correlation value is low. Pearson‟s and Spearman‟s correlation for simulated blur
is 0.748 and 0.744 respectively. Inclusion of real blur images for validation authenticates the
performance of this algorithm though the correlation is low.
Jing Zhang et al. [48] presented a kurtosis based quality metric for JPEG-2000 compressed
images. 1D kurtosis calculated in DCT domain is used as a quality indicator. Kurtosis represents
the deviation of the probability distribution from normal distribution and is measured as forth
central moment. Kurtosis increases as blurring increases. JPEG-2000 images from LIVE
database [12] are used for validation of algorithm. Pearson‟s and Spearman‟s correlation
coefficient is obtained as 0.922 and 0.915 respectively. This metric is designed only for JPEG-
2000 compressed images and uses blur as the sole parameter for quality evaluation.
Chaofeng Li et al. [49] proposed general regression neural network based approach for image
quality assessment. Phase congruency image [50], entropy of phase congruency image and
gradient of the distorted image are used as features. These features and Differential Mean
Opinion Score (DMOS) i.e., subjective image rating are fed to the neural network to obtain a
relationship. LIVE database [12] is used for validation purpose. Pearson‟s correlation,
Spearman‟s correlation coefficient and RMSE are obtained as 0.8374, 0.8268 and 8.7495
respectively. The correlation values for this metric are low as compared to other algorithms.
Ming-Jun Chen and Alan C Bovik [51] introduced a blur assessment metric based on Natural
Scene Statistics (NSS) model and wavelet decomposition. A Support Vector Machine (SVM)
[52] classifier is used to measure distance between image gradient statistics and NSS. Sum of
horizontal and vertical responses in the high band of wavelet decomposition is used to produce a
detail distortion map of image. Quality score is the function of distance given by SVM and
wavelet decomposition response. The algorithm is validated on LIVE database and the
Spearman‟s correlation coefficient is obtained as 0.9352. Pearson‟s correlation coefficient is not
used for benchmarking of this metric and distortion caused only due to blurring is considered for
quality assessment.
Ji Shen et al. [53] presented an image quality assessment algorithm for noisy, blurry, JPEG-2000
and JPEG compressed images. The algorithm is based on hybrid of curvelet, wavelet and cosine
transform. It uses the property of natural images which occupy well defined clusters in the
transformed space. Image characteristics are expressed by probability distribution of logarithm of
the magnitude of curvelet coefficients. Curvelet transform is replaced by wavelet transform and
DCT and the same procedure is applied to extract image characteristics. Pearson‟s correlation
coefficient obtained on LIVE database [12] is 0.921. Effect of distortions on statistics of three
different transforms is studied in this paper.
N. D. Narvekar and L. J. Karam [54] proposed a blur assessment metric for images. It is based
on Cumulative Probability of Blur Detection (CPBD). In this method, the image is first divided
into 64x64 size patches, depending on edge information, the block is classified as edge or non-
edge block. Probability of detecting blur in edge pixel is measured as presented in [28].
Normalized histogram of blur detection probabilities is obtained which provides the probability
density function. The algorithm is validated on LIVE database [12], TID2008 database [40],
IRCCyN/IVC [25] and MICT [55] database. Images distorted by Gaussian blur and JPEG-2000
compression are considered for evaluation. Exhaustive validation results yield a set of correlation
coefficients, RMSE, Outlier‟s Ratio is presented in Table 1. As this metric estimates presence of
blur, it cannot be used for images with other distortion types.
Jing Zhang et al. [56] used structural activity for image quality assessment. Structural activity is
the property which changes predictably either in spatial or frequency domain, whose variations
can imply that there is a change in structural information of an image. Direction spread is used as
a local structural activity indicator and a multi-stage median filter based approach is used for
structural strength estimation. LIVE database [12] is used for validation purpose. Pearson‟s and
Spearman‟s correlation coefficient is obtained as 0.9315 and 0.9217 respectively.
A. K. Moorthy and A. C. Bovik [57] proposed Distortion Identification-based Image Verity and
INtegrity Evaluation (DIIVINE) index. DIIVINE is based on the hypothesis that statistical
properties of images change in presence of distortions. In this algorithm, the distorted image is
subjected to Wavelet decomposition to obtain band pass responses. The wavelet decomposition
employed a steerable pyramid [58] over two scales (1,2) and six orientations(0°, 30°, 60°, 90°,
120°, 150°). The extracted sub-band coefficients are then used to extract statistical features. The
algorithm is validated on LIVE database [12] and TID2008 dataset [40]. Pearson‟s correlation,
Spearman‟s correlation coefficient and RMSE on LIVE database are obtained as 0.917, 0.916
and 10.90 respectively. Spearman‟s correlation coefficient on TID2008 dataset is 0.889.
Spearman correlation coefficient on LIVE and TID2008 datasets are close to 0.9 indicating the
consistency of this index.
Sangwoo Lee and Sang Ju Park [59] presented a method to measure strength of blocking artifacts
in block coded images. It is observed that high frequency components are present at boundaries
of image blocks which have blocking artifacts. This blocking artifacts cause pixel luminance to
change abruptly at block boundaries. In this method, pixel luminance change across block
boundaries is used as an indicator of blockiness. JPEG coded images from LIVE database [12]
are used for validation. Pearson‟s and Spearman‟s correlation coefficient is obtained as 0.9847
and 0.9764 respectively. This metric is designed only for images distorted by blocking artifact
and is validated only on JPEG coded images from LIVE database.
Anish Mittal et al. [60] introduced a training free model for image quality assessment. Mean
Subtracted Contrast Normalized Coefficients (MSCN) are used to form a feature vector. Feature
vectors are calculated for all image patches of size 64x64 with 8x8 overlap. The obtained feature
vectors are clustered into 400 visual words using K-means clustering algorithm. Model fitting
procedure is used to attain latent quality factors in test image. Pearson‟s and Spearman‟s
correlation coefficient for LIVE database [12] is obtained as 0.79 and 0.80 respectively. The
correlation coefficients obtained for this metric are low as compared to other algorithms.
Peng Ye and David Doermann [61] proposed an image quality metric based on visual codebook.
Codebook means collection of specific features which specify quality of an image. It uses Gabor
filtering [62] in five frequencies and four orientations on image patches of size 8*8. Mean and
variance of filtered output is used to form a feature vector. K-means clustering is used to form
200 clusters corresponding to each image. Images from LIVE database [12] are utilized to
construct the codebook. Pearson‟s and Spearman‟s correlation coefficient for LIVE database [12]
is obtained as 0.928 and 0.930 respectively. Pearson‟s and Spearman‟s correlation coefficient for
CSIQ dataset [63] is obtained as 0.908 and 0.884 respectively. Correlation coefficients on LIVE
and CSIQ databases are nearly same which indicate the consistency of this metric.
M. A. Saad et al. [64] utilized a Natural Scene Statistics (NSS) model of Discrete Cosine
Transform (DCT) coefficients for evaluation of image quality. The method relies on Bayesian
inference model. Gaussian density model is applied to DCT coefficients of 5*5 patches. A
function of derived Gaussian model is obtained in the form of a simple Bayesian model that
predicts quality. Pearson‟s and Spearman‟s correlation coefficient for LIVE database [12] is
obtained as 0.9302 and 0.9306 respectively. This is a training based approach so the algorithm
performs well only on distortion on which it is trained.
Anish Mittal et al. [65] proposed Blind/Referenceless Image Spatial QUality Evaluator
(BRISQUE), as an image quality metric in spatial domain. This algorithm uses locally
normalized luminance [66], i.e., Mean Subtracted Contrast Normalized (MSCN) image (I') and is
calculated as shown in (9).
(9)
where „I‟ is the intensity image, „µ‟ and „σ‟ are the mean and standard deviation in a 3x3
window. It further applies a generalized Gaussian distribution to obtain distorted image statistics
which tend to show change in coefficients distribution in presence of distortion. Pearson‟s and
Spearman‟s correlation coefficient for LIVE database [12] is obtained as 0.9424 and 0.9395
respectively. Spearman‟s correlation coefficient for TID2008 dataset [40] is obtained as 0.896.
The algorithm is benchmarked on two databases with consistent results.
Anish Mittal et al [67] proposed Natural Image Quality Evaluator (NIQE), an image quality
metric in spatial domain based on Natural Scene Statistics (NSS) model. In this method, the
image is first subjected to local mean removal and divisive normalization. Local mean deviation
is then calculated. Image quality is estimated by the distance between multi-variate Gaussian fit
of NSS features of test images and that of natural images. Pearson‟s and Spearman‟s correlation
coefficient for LIVE database [12] is obtained as 0.9147 and 0.9135 respectively. It is training
based approach and can be applied only on distorted images for which it is trained.
Xiaotong Huang et al. [68] presented a local image variance based approach to measure block
homogeneity. An adaptive method is used to select image blocks for noise statistics estimation.
Variance of selected blocks is used as an estimate for noise statistics. LIVE database [12] images
are used for validation and the Spearman‟s correlation coefficient is obtained as 0.9616.
Pearson‟s correlation coefficient is not used to benchmark this metric.
Shuhong Jiao et al [69] introduced a low computation image quality assessment metric in spatial
domain. This method uses the Mean Subtracted Contrast Normalized (MSCN) coefficients [65]
for log-energy computation. Log energy of each block of size 96x96 is calculated as shown in
(10).
(10)
Variance of MSCN coefficients is used to measure local contrast Image is divided into 96x96
patches and then operated. Pearson‟s and Spearman‟s correlation coefficient for LIVE database
[12] is obtained as 0.8815 and 0.9027 respectively. As it is a spatial domain metric, its
computation time is less.
Ming-Jun Chen et al. [70] proposed and image quality assessment model for static stereoscopic
images. In this method, a disparity map is generated from stereo image pair and multi-scale
Gabor filter responses are obtained. A cyclopean image is integrated from stereo image pair,
disparity map and Gabor filter responses. 2D features are extracted from cyclopean image and
3D features are extracted from the estimated disparity map. These features are then fed to a
training model used to predict image quality. LIVE 3D database [71] is used to evaluate
performance of this method. Pearson‟s, Spearman‟s correlation coefficients and RMSE for LIVE
3D database [71] is obtained as 0.895, 0.880 and 7.247 respectively.
Serir et al. [72] presented a blur metric for image quality assessment. The metric uses
Multiplicative Multi-resolution Decomposition (MMD) [73] in which the singularities are
characterized by ratio of polyphase components. MMD is similar to sub-band decomposition
using filter banks. It is a training algorithm and learns from parameters obtained from MMD.
The algorithm is validated using LIVE database [12], TID2008 dataset [40] and IRCCyN/IVC
database [25]. The details of results are shown in Table 1. This metric is benchmarked on three
datasets and the performance indicates its consistency.
Table 1: Various algorithms for No-Reference Image Quality Assessment
Ref.
Parameters used
PCC
SCC
RMSE
OR
Database
Distortion
tackled
[11]
Edge width
73%
81%
-
-
LIVE JPEG-
2000
Blur,
JPEG-2000
[13]
HVS model
-
-
-
-
Lena image
Multiple
[14]
Wavelet transform
0.91
-
8.54
-
LIVE JPEG-
2000
JPEG-2000
[15]
HVS model
-
-
0.57
-
LIVE JPEG
JPEG
[17]
PSF for blur
-
-
-
-
Blur
[18]
Renyi entropy
-
-
-
-
LIVE
Multiple
[20]
Standard deviation of
pixels
0.93
0.99
-
0.03
LIVE JPEG-
2000
JPEG-2000
[22]
NSS of DCT coefficients
0.97
0.97
-
-
LIVE
Multiple
[24]
Luminance and texture
features
0.96
0.96
-
-
IRCCyN/IVC
JPEG
0.96
0.90
-
-
LIVE JPEG
[26]
HVS features
-
-
0.7
-
LIVE JPEG
JPEG
[28]
JNB
0.93
0.93
0.46
0.41
LIVE G.Blur
Blur
0.88
0.87
0.39
0.47
LIVE JPEG-
2000
[30]
Radon transform
-
-
-
-
13 real blur
images
Blur
[31]
Edges
0.91
-
-
-
LIVE JPEG
Blockiness
[32]
Contourlet transform
0.82
0.73
14.10
-
LIVE
Multiple
[33]
Edge map
0.86
0.85
-
-
Kodak lossless
true color
image suite
Ringing
[34]
Wavelet transform
0.82
0.81
15.62
LIVE
Multiple
[37]
DCT
-
0.79
-
-
LIVE
Multiple
[38]
Pixel activity
0.92
0.91
6.04
0.03
LIVE JPEG-
2000
JPEG-2000
[39]
JND for blur,
Local variance for ringing
0.94
0.91
7.5
1.32
%
LIVE JPEG-
2000
JPEG-2000
0.91
0.90
0.81
-
LIVE G.Blur
0.92
0.94
8.2
-
TID2008
[41]
Fourier transform and PSF
-
-
-
-
G.Noise
corrupted
images
Blur,
G.Noise
[43]
Wavelet transform and
neural network
0.56
0.56
-
-
Real blur
Blur
0.74
0.74
-
-
Simulated blur
[48]
1D DCT Kurtosis
0.92
0.91
9.90
0.62
LIVE JPEG-
JPEG-2000
2000
[49]
Phase congruency
0.83
0.82
8.74
-
LIVE
Multiple
[51]
NSS and wavelet
transform
0.93
-
LIVE
Multiple
[53]
Hybrid of curvelet,
wavelet and DCT
0.92
-
-
-
LIVE
Multiple
[54]
Cumulative probability of
blur detection (CPBD)
0.91
0.94
8.98
0.16
LIVE G.Blur
Blur
0.88
0.88
11.42
0.32
LIVE JPEG-
2000
0.83
0.84
0.64
-
TID2008
G.Blur
0.92
0.92
0.74
-
TID2008
JPEG-2000
0.88
0.84
0.66
-
IRCCyN/IVC
G.Blur
0.77
0.78
0.86
-
IRCCyN/IVC
JPEG-2000
0.78
0.78
0.82
0.14
MICT JPEG-
2000
[56]
Median filter
0.93
0.92
-
-
LIVE
Multiple
[57]
Wavelet transform
0.91
0.91
10.9
-
LIVE
Multiple
-
0.88
-
-
TID2008
[59]
Luminance change across
0.98
0.97
-
-
LIVE JPEG-
Blockiness
blocks
2000
[60]
Locally normalized
luminance
0.79
0.8
-
-
LIVE
Multiple
[61]
Gabor filtering
0.92
0.93
-
-
LIVE
Multiple
0.90
0.88
-
-
CSIQ
[64]
NSS model of DCT
coefficients
0.93
0.93
-
-
LIVE
Multiple
[65]
Divisive normalization
0.94
0.93
-
-
LIVE
Multiple
-
0.89
-
-
TID2008
[67]
NSS in spatial domain
0.91
0.91
-
-
LIVE
Multiple
[68]
Local variance
-
0.96
-
-
LIVE
Multiple
[69]
Mean subtracted contrast
normalized coefficients
0.88
0.90
-
-
LIVE
Multiple
[70]
Gabor filter
0.89
0.88
7.24
-
LIVE 3D
Natural
Stereo
pairs
[72]
Multiplicative multi-
resolution decomposition
-
0.92
-
-
LIVE
Blur
0.90
0.89
-
-
TID2008
0.91
0.89
-
-
IVC
4. Image Quality Assessment databases
Results of a no-reference image quality assessment algorithm should agree well with human
perception of image quality. To analyze the performance of NR-IQA, standard image databases
are used. These databases are provided with subjective ratings (quality score) of images which
are obtained under standard test conditions. The quality score corresponding to each image is
also called Mean Opinion Score (MOS) or Differential Mean Opinion Score (DMOS).
Correlation of NR-IQA algorithm‟s output and DMOS indicates the performance of that
particular algorithm. Details of databases used for validation are presented in Table 2.
Information about other databases can be found in [74]. Pearson‟s correlation and Spearman‟s
correlation coefficient are mainly used to compare the performance of algorithm. Range of both
the correlation metrics is from „-1‟ to „+1‟ and value close to „-1‟ or „+1‟ indicates high
correlation in negative or positive direction respectively.
Table 2: Database for image quality assessment
Database
Distortions covered
Total images
Laboratory for Image &
Video Engineering (LIVE)
[12]
Reference images
29
JPEG
169
JPEG-2000
175
Gaussian Blur
145
White Noise
145
Fast fading
145
CSIQ [63]
Reference images
30
JPEG
150
JPEG-2000
150
Global contrast decrements
150
additive pink Gaussian noise
150
Gaussian blurring
150
TID 2008 [40]
17 types of distortions
25 reference +
1700 distorted
IRCCyN/IVC [25]
JPEG, JPEG-2000, LAR (Locally Adaptive
Resolution) compression and blur
10 reference + 235
distorted
MICT [55]
JPEG
98
JPEG-2000
98
5. Discussion:
An overview of discussed algorithms is provided in Table 1. Correlation coefficient values
corresponding to the algorithms are close to „1‟ but none has correlation equal to „1‟. Also
majority of algorithms target only specific types of distortions. Distortion caused by block
coding techniques and blur is a main area of focus so far. This shows the distortion specific
nature of algorithms. An ideal algorithm or may be a set of algorithms is expected to perform
equally well for all types of distortions. Algorithms are benchmarked using different databases
which mainly contain images degraded by limited number of distortions and each image contains
only one type of distortion. This may not be the situation in real life. Practical images are
frequently degraded by many types of noise, degradations and distortions. Correlation value
yield by an algorithm varies with the type of database used for validation and also from image to
image. So an algorithm needs to be benchmarked on all major image quality assessment
databases mentioned in Table 2. The best algorithm will yield high correlation coefficient values
on all the databases
6. Conclusion
To evaluate quality of an image in absence of a reference image is comparatively a complex task
that demands intelligence. Knowledge of statistics of natural images, response of human visual
system, the effect of distortions on images and also knowledge of the content of images is
required for designing better image quality assessment algorithm. Designing of such a system is
challenging as human knowledge, perception and inference is difficult to model mathematically.
Researchers are trying to mimic human perception using training approaches and statistical
parameters. But the techniques still have limitations as the performance of an algorithm is
limited by modelling or training. This paper reviews the methods developed for image quality
assessment in the absence of reference image with regard to the methods used, distortion nature,
testing and validating databases. The algorithms are benchmarked using subjective scores
provided with the corresponding database. Information about the databases mainly used for
quality assessment and the types of degraded images in the database is summarized. This survey
is expected to provide perspective researchers a complete overview of existing NR-IQA
algorithms at one place. Fortunately, in many specific practical applications, the statistical model
of images, the types of noise or distortion available in the images and their statistical models are
available. However, NR-IQA of an image of which statistics are not available and the type of
degradation or distortion is totally unknown is really a challenging problem. In practical distorted
images, the distortions, degradation, noise of different type may simultaneously affect the image
quality. This makes the NR-IQA problem further challenging for study, analysis, research and
development.
References:
[1] Q. Huynh-Thu, M. Ghanbari, Scope of validity of PSNR in image/video quality assessment,
Electronics letters 44 (2008) 800-801.
[2] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Simoncelli, Image quality assessment: From error
visibility to structural similarity, IEEE Trans. Image Process. 13 (2004) 600-612.
[3] H.R. Sheikh, A.C. Bovik, Image information and visual quality, IEEE Trans. Image Process.
15 (2006) 430-444.
[4] M.J. Chen, A.C. Bovik, Fast structural similarity index algorithm, Journal of Real-Time
Image Processing 6 (2011) 281-287.
[5] H.R. Sheikh, A.C. Bovik, G. de Veciana, An information fidelity criterion for image quality
assessment using natural scene statistics, IEEE Trans. Image Process. 14 (2005) 2117-2128.
[6] Z. Wang, E.P. Simoncelli, A.C. Bovik, Multiscale structural similarity for image quality
assessment, in: Conference Record of the Thirty-Seventh Asilomar Conference on Signals,
Systems and Computers, 2004, Vol.2, pp. 1398-1402.
[7] C. Li, A.C. Bovik, Content-partitioned structural similarity index for image quality
assessment, Signal Processing: Image Communication 25 (2010) 517-526.
[8] Z. Wang, A.C. Bovik, Modern image quality assessment, Synthesis Lectures on Image,
Video, and Multimedia Processing, Morgan and Claypool Publishers, United States of America,
2006.
[9] G. Xinbo, L. Wen, T. Dacheng, L. Xuelong, Image Quality Assessment Based on Multiscale
Geometric Analysis, IEEE Trans. Image Process. 18 (2009) 1409-1423.
[10] T. Dacheng, L. Xuelong, L. Wen, G. Xinbo, Reduced-Reference IQA in Contourlet
Domain, IEEE Trans. Syst., Man, Cybern. B, Cybern. 39 (2009) 1623-1627.
[11] P. Marziliano, F. Dufaux, S. Winkler, T. Ebrahimi, Perceptual blur and ringing metrics:
application to JPEG2000, Signal Processing: Image Communication 19 (2004) 163-172.
[12] Z.W. H.R. Sheikh, L. Cormack and A.C. Bovik, LIVE Image Quality Assessment Database
Release 2, http://live.ece.utexas.edu/research/quality.
[13] J. Du, Y. Yu, S. Xie, A new image quality assessment based on HVS, Journal of Electronics
(China) 22 (2005) 315-320.
[14] H.R. Sheikh, A.C. Bovik, L. Cormack, No-reference quality assessment using natural scene
statistics: JPEG2000, IEEE Trans. Image Process. 14 (2005) 1918-1927.
[15] R. Venkatesh Babu, S. Suresh, A. Perkis, No-reference JPEG-image quality assessment
using GAP-RBF, Signal Processing, 87 (2007) 1493-1503.
[16] G.B. Huang, P. Saratchandran, N. Sundararajan, An efficient sequential learning algorithm
for growing and pruning RBF (GAP-RBF) networks, IEEE Trans. Syst., Man, Cybern. B,
Cybern. 34 (2004) 2284-2292.
[17] I.v.Z. Marais, W.H. Steyn, Robust defocus blur identification in the context of blind image
quality assessment, Signal Processing: Image Communication 22 (2007) 833-844.
[18] S. Gabarda, G. Cristóbal, Blind image quality assessment through anisotropy, JOSA A 24
(2007) B42-B51.
[19] W.J. Williams, M.L. Brown, A.O. Hero III, Uncertainty, information, and time-frequency
distributions, in: San Diego,'91, San Diego, CA. International Society for Optics and Photonics,
1991. pp. 144-156.
[20] Z.M.P. Sazzad, Y. Kawayoke, Y. Horita, No reference image quality assessment for
JPEG2000 based on spatial features, Signal Processing: Image Communication 23 (2008) 257-
268.
[21] Z.M.P. Sazzad, Y. Kawayoke, Y. Horita, Spatial Features Based No Reference Image
Quality Assessment for JPEG2000, in: ICIP, IEEE International Conference on Image
Processing, 2007, Vol. 3, pp. 517- 520.
[22] T. Brandão, M.P. Queluz, No-reference image quality assessment based on DCT domain
statistics, Signal Processing 88 (2008) 822-833.
[23] E.Y. Lam, J.W. Goodman, A mathematical analysis of the DCT coefficient distributions for
images, IEEE Trans. Image Process. 9 (2000) 1661-1666.
[24] G. Zhai, W. Zhang, X. Yang, W. Lin, Y. Xu, No-reference noticeable blockiness estimation
in images, Signal Processing: Image Communication 23 (2008) 417-432.
[25] P.L. Callet, F. Autrusseau, Subjective quality assessment IRCCyN/IVC database,
http://www.irccyn.ec-nantes.fr/ivcdb/.
[26] S. Suresh, R. Venkatesh Babu, H.J. Kim, No-reference image quality assessment using
modified extreme learning machine classifier, Applied Soft Computing 9 (2009) 541-552.
[27] G.B. Huang, Q.Y. Zhu, C.K. Siew, Extreme learning machine: Theory and applications,
Neurocomputing 70 (2006) 489-501.
[28] R. Ferzli, L.J. Karam, A no-reference objective image sharpness metric based on the notion
of just noticeable blur (JNB), IEEE Trans. Image Process. 18 (2009) 717-728.
[29] N. Jayant, J. Johnston, R. Safranek, Signal compression based on models of human
perception, Proceedings of the IEEE 81 (1993) 1385-1422.
[30] S. Wu, W. Lin, S. Xie, Z. Lu, E.P. Ong, S. Yao, Blind blur assessment for vision-based
applications, Journal of Visual Communication and Image Representation 20 (2009) 231-241.
[31] S. Suthaharan, No-reference visually significant blocking artifact metric for natural scene
images, Signal Processing 89 (2009) 1647-1652.
[32] W. Lu, K. Zeng, D. Tao, Y. Yuan, X. Gao, No-reference image quality assessment in
contourlet domain, Neurocomputing 73 (2010) 784-794.
[33] H. Liu, N. Klomp, I. Heynderickx, A no-reference metric for perceived ringing artifacts in
images, IEEE Transactions on Circuits and Systems for Video Technology 20 (2010) 529-539.
[34] A.K. Moorthy, A.C. Bovik, A two-step framework for constructing blind image quality
indices, Signal Processing Letters, IEEE 17 (2010) 513-516.
[35] I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE
Transactions on Information Theory 36 (1990) 961-1005.
[36] V. Vapnik, The nature of statistical learning theory, Springer, New York Inc., 2000.
[37] M.A. Saad, A.C. Bovik, C. Charrier, A DCT statistics-based blind image quality index,
Signal Processing Letters, IEEE 17 (2010) 583-586.
[38] J. Zhang, T.M. Le, A new no-reference quality metric for JPEG2000 images, IEEE
Transactions on Consumer Electronics 56 (2010) 743-750.
[39] L. Liang, S. Wang, J. Chen, S. Ma, D. Zhao, W. Gao, No-reference perceptual image quality
metric using gradient profiles for JPEG2000, Signal Processing: Image Communication 25
(2010) 502-516.
[40] N. Ponomarenko, V. Lukin, A. Zelensky, K. Egiazarian, M. Carli, F. Battisti, TID2008 - A
Database for Evaluation of Full-Reference Visual Quality Assessment Metrics, Advances of
Modern Radioelectronics 10(2009), 30-45
[41] E. Cohen, Y. Yitzhaky, No-reference assessment of blur and noise impacts on image
quality, Signal, image and video processing 4 (2010) 289-302.
[42] Y. Yitzhaky, Yitzhak Yitzhaky‟s Home Page at Ben-Gurion University, Israel.
http://www.ee.bgu.ac.il/~itzik/nr_quality/.
[43] A. Ciancio, A.T. da Costa, A.B. da Silva, A. Said, R. Samadani, P. Obrador, No-reference
blur assessment of digital pictures based on multifeature classifiers, IEEE Trans. Image Process.
20 (2011) 64-75.
[44] X. Marichal, M. Wei-Ying, Z. Hongjiang, Blur determination in the compressed domain
using DCT information, in: ICIP, IEEE International Conference on Image Processing, 1999,
Vol. 382 pp. 386-390.
[45] T. Hanghang, L. Mingjing, Z. Hongjiang, Z. Changshui, Blur detection for digital images
using wavelet transform, in: ICME, IEEE International Conference on Multimedia and Expo,
2004, Vol. 11 pp. 17-20.
[46] R. Ferzli, L.J. Karam, Human Visual System Based No-Reference Objective Image
Sharpness Metric, in: IEEE International Conference on Image Processing, 2006, pp. 2949-2952.
[47] J. Mannos, D.J. Sakrison, The effects of a visual fidelity criterion of the encoding of images,
IEEE Transactions on Information Theory 20 (1974) 525-536.
[48] J. Zhang, S.H. Ong, T.M. Le, Kurtosis-based no-reference quality assessment of JPEG2000
images, Signal Processing: Image Communication 26 (2011) 13-23.
[49] C. Li, A.C. Bovik, X. Wu, Blind image quality assessment using a general regression neural
network, IEEE Transactions on Neural Networks 22 (2011) 793-799.
[50] M.C. Morrone, R.A. Owens, Feature detection from local energy, Pattern Recognition
Letters 6 (1987) 303-313.
[51] M.J. Chen, A.C. Bovik, No-reference image blur assessment using multiscale gradient,
EURASIP Journal on Image and Video Processing 1 (2011), 1-11.
[52] C.C Chang, C.J. Lin, LIBSVM : a library for support vector machines, ACM Transactions
on Intelligent Systems and Technology, 2 (2011). Software available at
http://www.csie.ntu.edu.tw/~cjlin/libsvm
[53] J. Shen, Q. Li, G. Erlebacher, Hybrid no-reference natural image quality assessment of
noisy, blurry, JPEG2000, and JPEG images, IEEE Trans. Image Process. 20 (2011) 2089-2098.
[54] N.D. Narvekar, L.J. Karam, A no-reference image blur metric based on the cumulative
probability of blur detection (CPBD), IEEE Trans. Image Process. 20 (2011) 2678-2683.
[55] Z. M. P. Sazzad, Y. Kawayoke, and Y. Horita, MICT Image quality evaluation database,
http://mict.eng.u-toyama.ac.jp/database_toyama/.
[56] J. Zhang, T.M. Le, S. Ong, T.Q. Nguyen, No-reference image quality assessment using
structural activity, Signal Processing 91 (2011) 2575-2588.
[57] A.K. Moorthy, A.C. Bovik, Blind image quality assessment: From natural scene statistics to
perceptual quality, IEEE Trans. Image Process. 20 (2011) 3350-3364.
[58] E.P. Simoncelli, W.T. Freeman, E.H. Adelson, D.J. Heeger, Shiftable multiscale transforms,
IEEE Transactions on Information Theory 38 (1992) 587-607.
[59] S. Lee, S.J. Park, A new image quality assessment method to detect and measure strength of
blocking artifacts, Signal Processing: Image Communication 27 (2012) 31-38.
[60] A. Mittal, G.S. Muralidhar, J. Ghosh, A.C. Bovik, Blind image quality assessment without
human training using latent quality factors, IEEE Signal Processing Letters 19 (2012) 75-78.
[61] P. Ye, D. Doermann, No-reference image quality assessment using visual codebooks, IEEE
Trans. Image Process. 21 (2012) 3129-3138.
[62] V. Kyrki, J.K. Kamarainen, H. Kälviäinen, Simple Gabor feature space for invariant object
recognition, Pattern Recognition Letters 25 (2004) 311-318.
[63] E.C. Larson, D.M. Chandler, The CSIQ image database, http://vision.okstate.edu/?loc=csiq.
[64] M.A. Saad, A.C. Bovik, C. Charrier, Blind image quality assessment: A natural scene
statistics approach in the DCT domain, IEEE Trans. Image Process. 21 (2012) 3339-3352.
[65] A. Mittal, A.K. Moorthy, A.C. Bovik, No-reference image quality assessment in the spatial
domain, IEEE Trans. Image Process. 21 (2012) 4695-4708.
[66] D.L. Ruderman, The statistics of natural images, Network: Computation in Neural Systems
5 (1994) 517-548.
[67] A. Mittal, R. Soundararajan, A.C. Bovik, Making a „completely blind‟image quality
analyzer, IEEE Signal Processing Letters 20.3 (2013), 209-212.
[68] X. Huang, L. Chen, J. Tian, X. Zhang, X. Fu, Blind noisy image quality assessment using
block homogeneity, Computers & Electrical Engineering (2013).
[69] S. Jiao, H. Qi, W. Lin, W. Shen, Fast and efficient blind image quality index in spatial
domain, Electronics Letters 49 (2013) 1137-1138.
[70] M.J. Chen, L.K. Cormack, A.C. Bovik, No-Reference Quality Assessment of Natural
Stereopairs, IEEE Trans. Image Process. 22 (2013) 3379-3391.
[71] A.K. Moorthy, C.C. Su, A. Mittal, A.C. Bovik, Subjective evaluation of stereoscopic image
quality, Signal Processing: Image Communication 28 (2012) 870-883.
[72] A. Serir, A. Beghdadi, F. Kerouh, No-Reference Blur Image Quality Measure Based on
Multiplicative Mutiresolution Decomposition, Journal of Visual Communication and Image
Representation 24 (2013) 911-925.
[73] A. Serir, A. Belouchrani, Multiplicative multiresolution decomposition for 2D signals:
application to speckle reduction in SAR images, in: ICIP, IEEE International Conference on
Image Processing, 2004, Vol. 651 pp. 657-660.
[74] S. Winkler, Analysis of Public Image and Video Databases for Quality Assessment, IEEE
Journal of Selected Topics in Signal Processing 6 (2012) 616-625.