International Journal of Computer Applications (0975 – 8887)

Volume 163 – No 6, April 2017

A Brief Review on Blind Image Quality Evaluation

Methods
Sushilkumar N. Holambe Ulhas B. Shinde Priyanka M. Kshirsagar
Persuing PhD at Dean Persuing ME
Dr. B.A.M.U. Aurangabad Faculty of Engg. & Technology at Terna college of Engg.
Dr. B.A.M.U. Aurangabad Osmanabad

ABSTRACT for compression and communication of digital image and

Image Quality Assessment plays an important role in various video services over wired and wireless networks. Image
image processing applications. It is still an active area of quality methods can be categorized in two parts subjective
research. A great deal of effort has been made in recent years and objective. The subjective assessment of image is done on
to develop objective image quality metrics that correlate well the bases of subjective experiments. While objective image
with perceived human quality measurement or subjective quality assessment methods were mainly based on some
methods. Image quality assessment means estimating the mathematical measures. The past five years have
quality of an image and it is used for many image processing demonstrated and witnessed the tremendous and imminent
applications. Image quality can be measured in two ways, demands of visual quality assessment metrics in various
subjective and objective method. In Subjective image quality applications.
assessment the evaluation of quality by humans is obtained by
mean opinion score (MOS) method where in objective
2. IMAGE QUALITY ASSESSMENT
evaluation of quality is done by algorithms. It concerned with METHODS
how image is perceived by a viewer and gives his or her Objective quality assessment is a very complicated task, and
opinion on a particular image and judge quality of the even full-reference QA methods have had only limited
multimedia content. The human eyes extract structural success in making accurate quality predictions. Researchers
information from the viewing field, so the human visual therefore tend to break up the problem of NR QA into smaller,
system is highly adapted for this purpose. domain-specific problems by targeting a limited class of
artifacts-distortion-specific IQA.
General Terms
Security, Pattern Recognition, Algorithms etc

Keywords
Image quality assessment, objective & subjective method

1. INTRODUCTION
With the development of imaging and multimedia
technologies, visual information, recorded by images has
become the main source for knowledge acquisition. In the
process of visual information acquisition, processing,
transmission, and storage, some artifacts or noise may be
introduced to images which degrade the visual quality. In a
typical digital imaging system, the image is captured and
transformed into digital signal by the sensor .This raw digital
image signal is then processed to reduce the noise and is
compressed for storage or transmission. When the image is Fig 1: Block diagram of Image Quality Assessment.
finally displayed on the screen to the end user, it might not be The most common being the blocking artifact, which is
same as the original version because it has been exposed to usually the result of block-based compression algorithms
various kinds of distortions. running at low bit rates. At live we have conducted research
The sources of distortion could be ranged from motion into NR QA for blocking distortion as well as pioneering
blurring, Gaussian noise, sensor inadequacy, compression, research into NR measurement of distortion introduced by
error during transmission or the combination of many factors. Wavelet based compression algorithms based on Natural
To improve the performance of visual information acquisition, Scene Statistics modeling. Following are the methods of
transmission, processing, and storage systems, it is essential to image quality assessment.
assess visual qualities of the images; so that it can maintain,
control and possibly enhance the quality of the image before
2.1 Natural Image Quality Evaluator
storage or transmission. The objective of image quality (NIQE)
assessment is to provide computational models to measure the Natural Image Quality Evaluator (NIQE) blind image quality
perceptual quality of a given image. Recently, a number of assessment (IQA) is a completely blind image quality
techniques have been designed to evaluate the quality of analyzer that uses only measurable deviations from statistical
images and videos. regularities observed in natural images, without training on
human-rated distorted images, and, indeed without any
The accurate prediction of quality from an end-user
exposure to distorted images. However, all current state-of-
perspective has received increased attention with the growing

24
International Journal of Computer Applications (0975 – 8887)
Volume 163 – No 6, April 2017

the-art general purpose no reference (NR) IQA algorithms The parameters of the GGD (α,β), can be reliably estimated
require knowledge about anticipated distortions in the form of using the moment-matching based approach proposed. The
training examples and corresponding human opinion scores. signs of the transformed image coefficients have been
observed to follow a fairly regular structure. However,
distortions disturb this correlation structure. This deviation
can be captured by analyzing the sample distribution of the
products of pairs of adjacent coefficients computed along
horizontal, vertical and diagonal orientation.
These natural scene statistic features are derived from a
collection of natural, undistorted images. Experimental results
show that the new index delivers performance comparable to
top performing NR IQA models that require training on large
databases of human opinions of distorted images. This method
uses only two types of the NSS features. It uses a single
global MVG model to describe the test image

2.2 BLind Image Integrity Notator using

DCT-Statistics (BLIINDS)
BLIINDS is one of the efficient, general-purpose, non-
distortion specific, blind that is no-reference image quality
assessment (NR-IQA) algorithm that uses natural scene
statistics models of discrete cosine transform (DCT)
coefficients to perform distortion-agnostic NR IQA. We
derive a generalized NSS-based model of local DCT
coefficients, and the model parameters are transform into
features suitable for perceptual image quality score prediction.
The statistics of the DCT features vary in a natural and
predictable manner as the image quality changes. We will
refer to undistorted images captured by imaging devices that
sense radiation from the visible spectrum as natural scenes,
and statistical models built for undistorted natural scenes as
NSS models. Deviations from NSS models, caused by the
introduction of distortions to images, can be used to predict
Fig 2: Patch Selection the perceptual quality of the image.
It depends on the construction of a quality aware collection of The model-based NSS-IQA approach developed here is a
statistical features that are based on a simple and successful process of feature extraction from the image, followed by
space domain of natural scene statistic (NSS) model. We use a statistical modeling of the extracted features. Purely NSS-
simple device to preferentially select from amongst a based IQA approaches require the development of a distance
collection of natural patches those that are richest in measure between a given distorted test image and the NSS
information and less likely to have been subjected to a model. This leads to the question of what constitutes
limiting distortion. This subset of patches is then used to appropriate and perceptually meaningful distance measures
construct a model of the statistics of natural image between distorted image features and NSS models. The
patches.The variance field has been largely ignored in the past Kullback– Leibler divergence as well as other distance
in NSS based image analysis, but it is a rich source of measures have been used for this purpose, but no perceptual
structural image information that can be used to quantify local justification has been provided for its use.
image sharpness. Letting the P×P sized patches be indexed b
Our approach relies on the IQA algorithm learning how the
=1,2,3,….B, a direct approach is to compute the average local
NSS model parameters vary across different perceptual levels
deviation field of each patch indexed b:
of image distortion. The algorithm is trained using features
derived directly from a generalized parametric statistical
model of natural image DCT coefficients against various
𝛿 𝑏 = 𝜎(𝑖, 𝑗) perceptual levels of image distortion. The learning model is
𝑖,𝑗 𝜖𝑝𝑎𝑡𝑐 ℎ𝑏
then used to predict perceptual image quality scores. Unlike
Given a collection of natural image patches selected as above, much of the prior work on image/video quality assessment
their statistics are characterized by „quality aware‟ NSS (QA) , we make little direct use of specific perceptual models
features computed from each selected patch [3]. Prior studies such as area V1 cortical decompositions, masking and motion
of NSS based image quality have shown that the generalized perception .Yet we consider our approach as perceptually
relevant since the NSS models reflect statistical properties of
Gaussian distribution effectively captures the behavior of the
the world that drive perceptual functions of the HVS. This is a
coefficients of natural and distorted versions of them. The
consequence of the belief that the HVS is adapted to the
generalized Gaussian distribution (GGD) with zero mean is
statistics of its visual natural environment. In other words,
given by:
models of natural scenes embody characteristics of the HVS,
𝛼 𝑥 which is hypothesized to be evolutionally adapted to models
𝑓 𝑥; 𝛼, 𝛽 = (−( )𝛼 )
exp⁡ conforming to natural scenes. HVS characteristics that are
2𝛽𝛾 1/𝛼 𝛽 intrinsic to, or that can be incorporated into NSS models
include:

25
International Journal of Computer Applications (0975 – 8887)
Volume 163 – No 6, April 2017

1. Visual sensitivity to structural information 2) BLIINDS is non-distortion specific; while most NR-IQA
algorithms quantify a specific type of distortion, the features
2. Perceptual masking used in our algorithm are derived independently of the type of
3. Visual sensitivity to directional information distortion of the image and are effective across multiple
distortion types. Consequently, it can be deployed in a wide
4. Multiscale spatial visual processing and range of applications.
5. Intolerance to flagrantly visible visual distortions. 3) We propose a novel model for the statistics of DCT
In the following sections we explain how one or more of these coefficients.
HVS properties are embedded in the model. The framework 4) Since the framework operates entirely in the DCT domain,
of the proposed approach is summarized in Fig. An image the one can take exploit the availability of platforms devised for
IQA “pipeline” is first subjected to local 2-D DCT coefficient the fast computation of DCT transforms.
computation.
5) The method requires minimal training, and relies on a
simple probabilistic model for quality score prediction. This
leads to further computational gains.
6) Finally, the method correlates highly with human visual
perception of quality and yields highly competitive
performance, even with respect to state-of-the-art FR-IQA
algorithms.

2.3 Distortion Identification-based Image

Verity and Integrity Evaluation (DIIVINE)
DIIVINE is a distortion-agnostic approach to blind IQA that
utilizes concepts from natural scene statistics (NSS) to not
only quantify the distortion and hence the quality of the
image, but also qualify the distortion type afflicting the image.
The Distortion Identification-based Image Verity and Integrity
Evaluation (DIIVINE) index utilizes a 2-stage framework for
blind IQA that first identifies the distortion afflicting the
image and then performs distortion-specific quality
assessment.
Our computational theory for distortion-agnostic blind IQA is
based on the regularity of natural scene statistics (NSS); for
example, it is known that the power spectrum of natural
scenes falloff as (approximately) 1/f^b , where f is frequency.
NSS models for natural images seek to capture and describe
the statistical relationships that are common across natural
(undistorted) images. Our hypothesis is that, the presence of
distortion in natural images alters the natural statistical
properties thereby rendering the image „un-natural‟. NR IQA
Fig 3: Overview of BLIINDS method can then be accomplished by quantifying this „un-naturalness‟
and relating it to perceived quality.
This stage of the pipeline consists of partitioning the image
into equally sized n x n blocks, henceforth referred to as local In order to extract statistics from distorted images we utilize
image patches, then computing a local 2-D DCT on each of the steerable pyramid decomposition. The steerable pyramid
the blocks. The coefficient extraction is performed locally in is an over complete wavelet transform that allows for
the spatial domain in accordance with the HVS‟s property of increased orientation selectivity. The choice of the wavelet
local spatial visual processing (i.e., in accordance with the fact transform was motivated by the fact that the scale-space
that the HVS processes the visual space locally). This DCT orientation decomposition that the wavelet transform performs
decomposition is accomplished across spatial scales. The mirrors models of spatial decomposition that occurs in area
second stage of the pipeline applies a generalized Gaussian V1 of the primary visual cortex. The steerable pyramid has
density model to each block of DCT coefficients, as well as been previously used for FR IQA as well as RR IQA with
for specific partitions within each DCT block. success. Note that we do not use the complex version of the
steerable pyramid as in, but that used in. Given an image
A generalized probabilistic model is applied to these features,
whose quality is to be assessed, the first step is to perform a
and are used to make probabilistic predictions of visual
wavelet decomposition using a steerable pyramid over 2
quality. We show that the method correlates highly with
scales and 6 orientations. We have found that an increased
human subjective judgements of quality. The contributions of
degree of orientation selectivity is beneficial for the purpose
this approach are as follows:
of QA - more so than selectivity over more than 2 scales. The
1) The proposed method inherits the advantages of the NSS choice of steerable filters was also motivated by its increased
approach to IQA. While the goal of IQA research is to orientation selectivity.
produce algorithms that accord with human visual perception
of quality, one can to some degree avoid modeling poorly
understood functions of the human visual system (HVS), and
resort to deriving models of the natural environment instead.

26
International Journal of Computer Applications (0975 – 8887)
Volume 163 – No 6, April 2017

performance improvements over the state-of-the-art. The

model proposed in extracts three sets of features based on the
statistics of natural images, distortion textures, and blur/noise;
three regression models are trained for each feature set and
finally a weighted combination of them is used to estimate the
image quality.

2.5 No-Reference Quality Assessment

algorithm for Block-Based compression
artifacts
Perhaps the most common distortion type that one comes
across in real-world applications is the distortion introduced
by lossy compression algorithms, such as JPEG (for images)
or MPEG/H.263 (for videos). These compression algorithms
are based on reduction of spatial redundancies using the
block-based Discrete Cosine Transform (DCT). When these
algorithms are constrained to increase the amount of
compression, a visible 'blocking' artifact can be seen.
Blocking resulting from DCT based compression algorithms
Fig 3: Distortion identification-based Image Verity and running at low bit rates has a very regular profile. It manifests
Integrity Evaluation (DIIVINE) index itself as an edge every 8 pixels (for the typical block-size of 8
x 8 pixels), oriented in the horizontal and vertical directions.
The Distortion Identification-based Image Verity and Integrity
The strength of the blocking artifact can be measured by
Evaluation (DIIVINE) – divines the quality of an image
estimating the strength of these block-edges. At LIVE, we
without any need for a reference or the benefit of distortion
have developed frequency domain algorithms for measuring
models, with such precision that its performance is
blocking artifact in images compressed by JPEG, with the
statistically indistinguishable from popular FR algorithms
algorithm having no information about the reference image.
such as the structural similarity index (SSIM). The DIIVINE
approach is distortion-agnostic, since it does not compute 2.6 No-Reference Quality Assessment for
distortion-specific indicators of quality, but utilizes an NSS-
based approach to qualify as well as quantify the distortion JPEG2000 Compressed Images using
afflicting the image. The approach is modular, in that it can Natural Scene Statistics
easily be extended beyond the pool of distortions considered Not all compression algorithms are block-based. Recent
here. This method assume that the distortion types in the test research in image and video coding algorithms has revealed
images are represented in the training dataset, which is, that a greater compression can be achieved for the same visual
however, not the case in many practical applications. quality if the block-based DCT approach is replaced by a
Discrete Wavelet Transform (DWT). JPEG2000 is a recent
2.4 Blind/Referenceless Image Spatial image compression standard that uses DWT for image
Quality Evaluator (BRISQUE) compression. However, DWT based algorithms also suffer
Blind/Referenceless Image Spatial Quality Evaluator from artifacts at low bit rates, specifically, from blurring and
(BRISQUE) is a natural scene statistic (NSS)-based ringing artifacts. Blurring and ringing artifacts are image
distortion-generic blind/no-reference (NR) image quality dependent, unlike the blocking artifact, whose spatial location
assessment (IQA) model which operates in the spatial domain. is predictable.
It does not compute distortion specific features such as This makes the task of quantifying distortion resulting from
ringing, blur or blocking, but instead uses scene statistics of DWT based compression algorithms (such as the JPEG2000)
locally normalized luminance coefficients to quantify possible much harder to quantify. At LIVE we have proposed a unique
losses of „naturalness‟ in the image due to the presence of and innovative solution to the problem. We propose to use
distortions, thereby leading to a holistic measure of Natural Scene Statistics models to quantify the departure of a
quality.The underlying features used derive from the distorted image from "expected" natural behavior.
empirical distribution of locally normalized luminances and
products of locally normalized luminance under a spatial 2.7 A No Reference Image Quality
natural scene statistic model. No transformation to another
coordinate frame (DCT, wavelet, etc) is required, Assessment by using a general regression
distinguishing it from prior no reference IQA approaches. neural network (GRNN)
Despite its simplicity, we are able to show that BRISQUE is The general regression neural network is a powerful
statistically better than the full-reference peak signal-to-noise regression tool that has a dynamic network structure. It is
ratio (PSNR) and the structural similarity index (SSIM) and based on established statistical principles, and asymptotically
highly competitive to all present-day distortion-generic NR converges with an increasing number of samples to the
IQA algorithms. BRISQUE has very low computational optimal regression surface. It extract image quality-related
complexity, making it well suited for real time applications. statistical features in both the spatial and frequency domains.
BRISQUE features may be used for distortion identification In the spatial domain, locally normalized pixels and adjacent
as well. pixel pairs were statistically modeled using log-derivative
statistics; and in the frequency domain, log-Gabor filters were
To illustrate a new practical application of BRISQUE, we used to extract the fine scales of the image. Based on the
describe how a non-blind image denoising algorithm can be observation that image local contrast features convey
augmented with BRISQUE in order to perform blind image important structural information that is related to image
denoising. Results show that BRISQUE augmentation leads to perceptual quality, a BIQA model utilizing the joint statistics

27
International Journal of Computer Applications (0975 – 8887)
Volume 163 – No 6, April 2017

of the local image gradient magnitudes and the Laplacian of [3] M. Saad, A. C. Bovik, and C. Charrier, “Blind image
Gaussian image responses is found. GRNN has been observed quality assessment: A natural scene statistics approach in
to yield better results than the back-propagation network or the DCT domain,” IEEE Trans. Image Process., vol. 21,
RBF (radial basis function) network in terms of prediction no. 8, pp. 3339–3352, 2012.
performance. The GRNN was implemented using the
MATLAB function new GRNN. The four perceptually [4] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli,
motivated features can be used as inputs to the GRNN: “Image quality assessment: From error visibility to
structural similarity,” IEEE Trans. Image Process., vol.
1) The mean value of the phase congruency image of 13, no. 4, pp. 600–612, Apr. 2004.
distorted image (MPC),
[5] S. Daly, “The visible difference predictor: An algorithm
2) The entropy of the phase congruency image of for the assessment of image fidelity,” in Digital images
distorted image (EPC), and human vision (A. B. Watson, ed.), pp. 179–206,
Cambridge, Massachusetts: The MIT Press, 1993.
3) The entropy of the distorted image (EDIS), and
[6] J. Lubin, “The use of psychophysical data and models in
4) The mean value of the gradient magnitude of the the analysis of display system performance,” in Digital
distorted image (MGDIS). images and human vision (A. B. Watson, ed.), pp. 163–
178, Cambridge, Massachusetts: The MIT Press, 1993.
3. ACKNOWLEDGMENTS
Our thanks to the experts who have contributed towards the [7] W. Xu and G. Hauske, “Picture quality evaluation based
research paper. on error segmentation,” Proc. SPIE, vol. 2308, pp. 1454–
1465, 1994.
4. REFERENCES
[1] A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference [8] W. Osberger, N. Bergmann, and A. Maeder, “An
image quality assessment in the spatial domain,” IEEE automatic image quality assessment technique
Trans. Image Process. (to appear), 2012. incorporating high level perceptual factors,” in Proc.
IEEE Int. Conf. Image Proc., pp. 414– 418, 1998.
[2] A. K. Moorthy and A. C. Bovik, “Blind image quality
assessment: From natural scene statistics to perceptual [9] A. Mittal, R. Soundararajan and A. C. Bovik, “ Making a
quality,” IEEE Trans. Image Process., vol. 20, no. 12, pp. Completely Blind Image QualityAnalyzer ”, IEEE Signal
3350–3364, 2011. processing Letters, pp. 209-212, vol. 22, no. 3, March
2013.

IJCATM : www.ijcaonline.org
28