UIJRT | United International Journal for Research & Technology | Volume 02, Issue 01, 2020 | ISSN: 2582-6832
Review of Techniques and Methods for Image
Deblurring
Sahibzada Muhammad Wahab
Department of Computer Systems Engineering,
University of Engineering and Technology (UET), Peshawar, Pakistan
sahibzada.wahab987@gmail.com
Abstract— Technological advancements in the
scientific world has witnessed its exponential growth
since the transformation from analogue to digital
devices. Computer Systems transitioned from
humongous sized poor performance gadgets to powerful
computational devices even surpassing the human
intelligence in most scenarios. Such advancements gave
birth to algorithmic approach to world most complex
problems. As algorithms, for their decision making
requires data, which in most cases is in the form of
images, those images for the sake for better processing
needs to be clear with sharp edges, which simply means
there shouldn’t be any kind of blur or noise effecting
those input images. Blur or noise are added to the input
images during the capturing process either because of
the natural scene lighting or the complexity of scene or
because of convolution of impulse response which is
called as blur kernel or point spread function(PSF).
Image processing field which deals with deblurring of
such images is called as image restoration. There are two
methodologies for dealing with such scenarios that are
blind image deblurring and reference based image
deblurring. This paper gives extensive review of
research done in both domains. Discussing in details the
approaches been used along with comparing the data to
analyze and specify the methodology that is well suited
for most blur scenarios.
Keywords— Image Processing, image restoration,
blind image deblurring, conventional image deblurring.
1. INTRODUCTION
Over the years our dependence on the algorithms has
increased significantly in every domain of our life
ranging from medical diagnosis to satellite imagery,
security to industrial processes.
Because additions of such degradations will affect the
readability of those images which in return will affect
the output of algorithmic process. To avoid such
degradations in the input images another field in the
image processing has emerged which is known as image
restoration.
In the field of image restoration, reversal of an image
from its degraded version to a clear, natural scene replica
version is a well-established issue. Because there isn’t
any single generic equation through which any blur
image can be deblurred.
The reason for its absence is the nature of degradation
which can be in quite a number of forms. In case of blur
affected images that blur kernel can one or combination
of many blur kernels, can be space variant or space
invariant, can be parametric or non-parametric.
So deducing single equation for all these possibilities is
a challenge that is yet to be completed. Like blur, noise
is also a degradation factor which can also be added due
to many reasons, like poor camera lens calibration, scene
lightening etc.
For tackling image degradation, different deblurring
techniques are introduced over the years. These
techniques are generically divided into two categories
depending upon the information we have related to the
blur image.
One is reference based image deblurring where either we
have information related to the Point Spread function
(PSF), which is impulse response convolved with the
original image at the time of capturing process or we
have true image which can be utilized by the
optimization algorithms as the stopping criteria when
required MSE between the deburred image and original
image is achieved or by the machine learning algorithms
for training.
In each of these domains one factor that is common is
the need of images either in shape of single image or
videos, which are fed to the systems to get the desired
output.
The second category is reference-less or blind image
deblurring where neither we have any information
related to the blur kernel nor we have the true image.
But for the images to be well interpreted they must be
clear which in the field of image processing means there
should not be any sort of degradation in the image either
in the form of blur or noise.
Review section will be divided in above-mentioned two
generic image restoration categories. It will also discuss
All rights are reserved by UIJRT.COM.
This paper will give extensive review of the research
done in the field of image restoration.
10
UIJRT | United International Journal for Research & Technology | Volume 02, Issue 01, 2020 | ISSN: 2582-6832
the data from different research approaches and will
analyze which restoration algorithm yields better results.
1.1 PROBLEM STATEMENT
Addition of blur or noise that is added to the images is
an undesirable phenomena especially in scenarios where
such degradations is highly unacceptable like that of
medical imagery where critical decisions depends on
such images or in satellite imagery where sending probe
or satellite back to last known locations for a single
image is financially not feasible or in CCTV footage
which is really important for security. For tackling such
scenarios, algorithms must be developed, which can
efficiently remove the unwanted blur and noise from the
image. But unfortunately these degradations cannot be
presented with a single mathematical equation which
results in the complexity of resolving such issues.
Various types of blur are discussed below:
Camera Shake Blur:
With handheld photography devices in trending,
inclusion of this kind of blur in images has increased
significantly because this blur is added to an image when
the camera is moved (translated or rotated or both)
during capturing process. Movement is not the only
reason, low lightening condition of the scene can also
cause the addition of camera shake blur. Below figure
depicts camera shake blur.
Motion blur:
At the image capturing instance both the camera device
and the object must be stationary, if either moves, it will
cause the addition of motion blur in the image. Along
with the movement, low exposure time is also the cause
of this kind of blur. Figure 3 show the simulated motion
blur effect.
Figure 3: Motion Blur [1]
Imperfect Focus:
External factors like temperature, wind, motion, camera
shake are not the only factors for the inclusion of blur in
an image, but sometimes poor calibrations of camera
lenses can also cause degradation in an image. This kind
of blur is known as imperfect focus blur.
Figure 1: Camera Shake Blur [1]
Atmospheric Turbulences:
In scenarios like satellite imagery, or other long range
image capturing systems, often temperature variation or
wind which causes light rays refractions causes
atmospheric turbulence blur. Figure 2 depicts
atmospheric turbulence effected simulated image.
Figure 2: Atmospheric Turbulence [1]
All rights are reserved by UIJRT.COM.
Figure 4: Imperfect Focus [1]
2. TECHNIQUES AND METHODS
As deblurring methodologies are grouped into two
major categories that are reference based image
deblurring and reference-less image deblurring. Hence
this section is divided such that research work done in
each category is explained separately.
2.1 Reference Less/Blind Image Deblurring
Finding blur kernel for the blurred image is always a
challenge in image restoration. Different techniques are
used for its retrieval, one technique among many, is the
use of optimization algorithms that are Genetic
algorithm, Ant colony algorithm and Particle swarm
11
UIJRT | United International Journal for Research & Technology | Volume 02, Issue 01, 2020 | ISSN: 2582-6832
optimization. These optimization algorithms converge
toward local and global minima in the most efficient
way which helps tremendously in finding the blur kernel
parameters
In [2] they have presented a technique where Genetic
algorithm optimization is utilized for deblurring of old
documents, documents that are hundreds of years old.
Digitizing such documents manually will take ages and
will be prone to errors. Speeding up this process requires
automatic system like OCR or document scanning etc.
But such methodologies are also liable to the addition of
blur or noise. The solution for such an issue is presented
by this research work where they have enhanced fitness
function calculation presented by [3], where number of
edges are summed up for fitness function which in case
of old documents will fail because old documents were
hand written where edges length might vary, so for
avoiding just ambiguities. They have enhanced this
system by calculating only those edges that are defined
by single pixel. Along with defining the fitness function
they also have experimented on median of the image
which shows that images preprocessed with median
filter does not yield better results. Figure 5 depicts the
output of above proposed methodology.
scene and fixing the scenario that the small camera
motion affected the image. There focus is on the
calculation of not only the non-uniform blur kernel but
also on the estimation of 6 Dof camera motion, because
by doing so, not only plane images but videos with depth
factor can be deblurred with ease while keeping the
video frames sequence in order to keep the video quality
sharper. For initial experimentation datasets like that of
KITTI and Middlebury were used where synthetic blur
were introduced on real life images but both of these
datasets lack the presence of depth map, for which
another dataset, TUM RGB-D was utilized. Figure 6
shows the outputted results.
(a)
(b)
Figure 6: (a) Blur Image and (b) Proposed
methodology [4]
In [5] for the construction of arbitrarily shaped PSF three
optimization algorithms were utilized for PSF quick
convergence, among the three algorithms there
experimented results depicted that the convergence rate
of Particle swarm optimization is much faster than that
of the other two that are Ant Colony Optimization and
Genetic algorithm. However in terms of programmatic
implementation Genetic algorithm leads the way as
compared to the other two algorithms. Outputs are
presented in figure below.
(a)
(b)
Figure 5: (a) Blur Image and (b) Proposed
methodology [2]
With the increasing popularity of smart phones,
handheld photography has been in the spotlight. But
with the reliability in the capturing of real-life moments,
absence of high-end hardware, causes the images to be
prone to blur, especially camera shake blur and motion
blur. The reason for camera shake blur as mentioned in
this paper, might be the complex structure of the reallife environment or the long exposure time of the
capturing process. Quite extensive research is being
going on in this domain of image restoration. The
solution proposed by [4] is to utilized the single blurring
image for the deblurring of whole video sequence while
assuming the availability of depth map of the observed
All rights are reserved by UIJRT.COM.
Figure 7: (a) Blur Image (b) Genetic Algorithm
Deblurring (c) Ant Colony Deblurring (d) Particle
Swarm Optimization Deblurring [5]
12
UIJRT | United International Journal for Research & Technology | Volume 02, Issue 01, 2020 | ISSN: 2582-6832
In [6] they have utilized the Genetic algorithm for the
local minima convergence for the PSF values which
were used by the deconvolution algorithm to generate
deblurred images, which after deconvolution were
analyzed for the whitenss of the residual image. The
outputted value determined the iteration status, either to
stop the process or to keep going. Same Genetic
algorithm is also utilized in [7] where PSF values
specifically for the Gaussian Blur parameters were
generated. The generated population of Gaussian blur
sigma values were utilized by the Weiner filter for
deconvolution. The resultant images were passed
through Kurtosis function for analysis, which
determined whether the image is deblurred or not.
Depending on the output of the Kurtosis function the
iteration status was determined. Results of above
methodology is presented in Figure 8.
In [9] methodology for bulk image deblurring is
presented where a unique approach of utilizing the basis
and corresponding coefficients of blurred images are
utilized. This process works amazingly for many blurred
images like in the case of video processing. Where
different frames are to be deblurred. Values calculated
for one image is utilized for the whole sequence. Thus
minimizing the computational workload significantly.
Figure 10 shows the results of proposed approach.
With the advancement in the mobile phone systems,
applications like gyroscope and accelerometer are
widely used in almost all the mobile devices. In [8] they
have proposed a techniques depending on above
mentioned features of today’s smart phones. At every
instant of using mobile devices, data related to
gyroscope and accelerometer is saved. Upon accessing
that data at the time of capturing of blurred image, it can
be utilized by Particle swarm optimization to generate
PSF values from it. Motion vector can be calculated by
the motion of cellphone during image capturing. Below
figure 9 outputted results depicts the efficiency of the
above proposed scheme.
(a)
(b)
Figure 10: (a) Blurred image and (b) Deblurred
Imaged [9]
Often images restored by the deblurring process losses
its detail. Which in cases like pattern recognition of
tissues patterns or satellite imagery, is unacceptable. To
address this problem, an effective approach is presented
in [10]. There proposed method involve the use of most
popular neural algorithm, BP neural network, which
upon the utilization of fractional form error functions as
its core error function and particle swarm optimization
for its optimization, yields better results than traditional
deblurring techniques like that of Wiener filter
deconvolution, constrained least square (CLS)
deblurring method and many more. Upon
experimentation they have proposed that deblurring
with CLS smoothed the sharp edges hence affecting the
important details while on the other had Weiner
deconvolution is better in removing blur but it creates
more jagged edges. 6 real life images were tested with
all the above-mentioned algorithms after training the BP
neural network. The results obtained from proposed
methodology retains the fine details in the deblurred
image along with the removal of blur from the image.
Figure 9: 1: True Image 2: Blind Deconvolution 3:
Wiener filter 4: Deblurred by Lucy-Richardson
method, 5: Regularized filter Method, and 6: proposed
methodology [8]
2.2 Reference Based Image Deblurring
In most of research work presented, spectrum of the
blurred picture is utilized for finding the angle parameter
of the blur kernel while Hough transform is utilized for
the direction in which blur kernel was convoluted with
the original image. But a different technique is
introduced in [11] which is based on equation given in
[12], where instead of above mentioned methodology
they have proposed a new scheme which utilized the
frequency domain representation of blurred image, and
they have utilized cestrum instead of spectrum for
Figure 8: Gaussian blur effected image with sigma=2
and resultant output image with proposed scheme with
calculated sigma=2.33 [7]
All rights are reserved by UIJRT.COM.
13
UIJRT | United International Journal for Research & Technology | Volume 02, Issue 01, 2020 | ISSN: 2582-6832
finding the angle while for direction they have utilized
Radon transform. But this algorithm fails in case of real
life blur kernel as those blurs are of irregular shapes.
Table 1: Proposed methodology on various images [11]
Figure 11: Deblurring with values from table 1 [11]
As in reference based image deblurring we either have
knowledge related to PSF or true image. In [13]
restoration of blurred images is performed by utilizing
Richardson-Lucy deconvolution algorithm instead of
wiener filter while having prior information related to
PSF along with focusing of Poisson noise instead of
Gaussian noise. For deconvolution different algorithms
are used, some of those popular algorithms are,
Algebraic deconvolution, Basis Pursuit deconvolution
and Richardson Lucy algorithm [14].
In [15] above mentioned three deconvolution algorithms
are compared with experimentations. There outputted
results depict that Basis Pursuit deconvolution yields
sharpest results of all and along with that it’s really
amazing when it comes to suppressing the ringing effect.
But when it comes to execution time Algebraic
deconvolution performs the best while its ringing
suppression quality was the worst.
Figure 12: (a) Blurred Image (b) Noisy Image (c) Blur
kernel (d) Richardson Lucy result (e)Algebraic
deconvolution result (f) BPD result [15]
Identifying the optimal solution or local minima for a
problem having different solution possibilities is
All rights are reserved by UIJRT.COM.
computationally expensive because to accept a specific
optimal solution the algorithm has to traverse through all
the possible solutions first, such problems lies in the
category of nonconvex optimization problem.
In [16] they focus on the objective of obtaining a blur
free image along with the calculation and identification
of the blur kernel which pushes it to the category of
nonconvex optimization.
They have constructed an algorithm based on proximal
alternating minimization (PAM). But for the algorithm
to generate better results an initial point must be given
to it that’s the reason their algorithm is semi-blind
deblurring algorithm.
Proposed algorithm works better then TV-RSTLS, no
matter the blur kernel, given the initial guess for both the
algorithms is same.
Blur kernels under considerations were Moffat kernel,
Gauss kernel and truncated non-symmetric kernel.
Quantitative results were based on the SNR, structural
similarity index and PSNR of the resultant images from
the both, the proposed methodology and the TV-RSTLS
algorithm.
(a)
(b)
Figure 13: (a) Blur Image and (b) Proposed
methodology [16]
In [17] an approach based on calculation of PSF by
finding threshold value is utilized where image is first
compressed to 255x255, then PSF values of extreme
high and low is applied on it. They have applied
Gaussian blur kernel.
The process is repeated till they find the threshold value.
For reducing the ringing effect, they have utilized
WEIGHT array method.
In [18] Genetic algorithm is utilized for the generation
of random population of PSF values.
From that population PSF values that yields better result
is utilized for the generation of another population.
This process continues until we get PSF values that
yields the best results.
14
UIJRT | United International Journal for Research & Technology | Volume 02, Issue 01, 2020 | ISSN: 2582-6832
3. CONCLUSION
Image deblurring has always presented a challenge for
the research community no matter if domain in spotlight
is reference based image deblurring or reference-less
image deblurring. But among the two, reference based
image deblurring is relatively easy, computationally less
expensive, yields better results than that of blind image
deblurring. The reason for such a difference is either
because of the availability of true image or because of
the availability related to point spread function.
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
REFERENCES
Wang, R. and D. Tao, Recent Progress in Image
Deblurring. 2014.
Deborah, H. and A.M. Arymurthy. Image
Enhancement and Image Restoration for Old
Document Image Using Genetic Algorithm. in 2010
Second International Conference on Advances in
Computing, Control, and Telecommunication
Technologies. 2010.
S. Hashemi, S.K., N. Noorozi, M. E. Moghaddam.
An image enhancement method based on genetic
algorithm. in IEEE International Conference on
Digital Image Processing. 2009.
Liyuan Pan, Y.D., Miaomiao, Single Image
Deblurring and Camera Motion Estimation with
Depth Map, in 2019 IEEE Winter Conference on
Applications of Computer Vision. 2019.
Waleed, M., A. Khan, and A. Khan. On the quick
convergence of PSF estimation for single image
blind deblurring. in 2017 9th International
Conference on Electronics, Computers and
Artificial Intelligence (ECAI). 2017.
Ali, P.S.M. and R. Begum, Improving Image
Deblurring Using Genetic Algorithm. International
Journal of Advanced Research in Electrical,
Electronics and Instrumentation Engineering, 2016.
Vol. 5(Issue 7).
Yin, H. and I. Hussain. Blind Source Separation and
Genetic Algorithm for Image Restoration. in 2006
International Conference on Advances in Space
Technologies. 2006.
Rajakaruna, N., et al., Image Deblurring for
Navigation Systems of Vision Impaired People
Using Sensor Fusion Data. 2014.
Li, W., et al., Fast non-blind image deblurring in
frequency domain based on matrix decomposition.
2016. 46-50.
Zhang, Y., et al. Image Restoration Based on
Robust Error Function and Particle Swarm
Optimization-BP Neural Network. in 2008 Fourth
International Conference on Natural Computation.
2008.
All rights are reserved by UIJRT.COM.
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Pathak, V., et al., An Efficient Algorithm For
Deblurring A Natural Image. International Journal
of Engineering Research and Applications (IJERA)
ISSN: 2248-9622, 2013. Vol. 3(Issue 3).
Lokhande, R., K.V. Arya, and P. Gupta,
Identification of parameters and restoration of
motion blurred images. Vol. 1. 2006. 301-305.
Jayapriya, P. and D.R.M. Chezhian, A Study on
Image Restoration and its Various Blind Image
Deconvolution Algorithms. International Journal of
Computer Science and Mobile Computing,
IJCSMC, 2013. Vol. 2(Issue. 10).
Richardson, W.H., Bayesian-Based Iterative
Method of Image Restoration*. Journal of the
Optical Society of America, 1972. 62(1): p. 55-59.
Das, R., A. Bajpai, and S. Venkatesan, Fast Nonblind Image Deblurring with Sparse Priors. 2017. p.
629-641.
Hong-Xia Doua, T.-Z.H., Xi-Le Zhaoa, Jie Huanga,
Jun Liub, Semi-blind image deblurring by a
proximal alternating minimization method with
convergence guarantees. ELSEVIER, 2020.
Bundela, A., A. Chourasiya, and U. Bhan Singh,
Restoration of Single Blur Image Using Blind
Deconvolution Method. Vol. 20. 2015. 72-76.
Bhavani, S.A., Implementation of Deblurring
Images Using Blind Deconvolution Technique.
International Journal of Science and Research
(IJSR), 2013.
15