International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023

RESEARCH ARTICLE OPEN ACCESS

Review of Single Blind Image Deblurring Techniques

Nikhil K. Pawanika [1], R. Srivaramangai [2]
[1]
Department of Information Technology, University of Mumbai - Mumbai
[2]
Department of Information Technology, University of Mumbai - Mumbai
ABSTRACT
Digital images are an inseparable part of our everyday personal lives where we capture any moment for sharing with others,
later viewing or in professional applications like remote sensing, medical imaging, surveillance, etc. The past decade saw an
unprecedented increase in the availability of affordable, portable, and highly functional digital multimedia devices such as
cameras, mobile-phones, digital recorders, etc. Due to this increased possibility of generating digital audio-visual content
without constraints of time, location, and network has been possible [1]. Although on one hand, there is a plethora of devices
that claim to deliver HD quality of images but that is true only for static images in an ideal setting, on the other hand there are
images that are often blurred due to various reasons that prevents us from exploring the necessary details making feature
extraction tasks such as object recognition, segmentation, alignment and stitching of images, 3D stereo reconstruction, self-
navigating robots or cars, etc. difficult [2]. This review paper focuses on the study of different techniques of single blind image
deblurring used for restoration & enhancement of blurred images.
Keywords: - Blur, blind deblurring, point spread function, priors, convolution neural networks, datasets, benchmarks, image
quality metrics.

I. INTRODUCTION
Digital images find their application in almost every aspect An easy way to comply with the conference paper
of life. Some areas where digital images are professionally formatting requirements is to use this document as a
used are surveillance systems, medical field, remote sensing, template and simply type your text into it.
pattern recognition, etc. as showed in Figure 1. Most of the
times professional digital images captured in ideal settings are
static or captured in automated fashion like in medical
imaging where the patient is stable. But in other cases, very
often digital images that are captured using handheld devices
like mobile phone or digital cameras contain blur which can
be due to various factors such as camera motion, motion of
a. Images from a b. MRI images of brain
object, angle of camera, light exposure, etc. Blurring of an surveillance system (medical imaging)
image is one of the major reasons for degradation of an image Courtesy: shutterstock Courtesy: thinkstock
[3][4]. Image blur is an undesired artefact in digital images
and prevents the accurate interpretation of scene or feature
extraction as it causes loss of necessary detail [5]. Examples
of types of blur identified in forensics include: Motion Blur,
Gaussian Blur, Out of Focus blur, Atmospheric blur, etc [6].
In Figure 2 below, the number of people & the license plates
could not be identified in the blurred images.
c. Ariel view of a farm d. Object Detection
(remote sensing) (pattern recognition)
This paper is organized as follows. Section 2 justifies the Courtesy: geoawesomeness Courtesy: jetbrains blog
need for writing this research paper. Section 3 defines the
terms on blurring its types, deblurring, its classification and Fig 1. Real Life Applications of Digital Images
image deblurring techniques. Section 4 presents the review of
literature in single blind image deblurring techniques. Section
5 presents the observations of the papers reviewed from II. RATIONALE
different perspectives. Section 6 presents the concluding Image deblurring, denoising & restoration has traditionally
remarks. been a challenging problem in image processing. When
images fail to achieve a desired level of quality with reference
to its clarity and/or features present in it, a process called
image restoration is applied to alleviate its quality. Poor
quality input images pose problems for high level image

ISSN: 2347-8578 www.ijcstjournal.org Page 129

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
processing task such as image classification, object Deblurring is the restoration of images, a process of
recognition, and scene understanding. Image restoration has a recovering an original image from the degraded image which
rich literature where different techniques have been developed has been blurred due to various factors mentioned above. The
over time with promising results but the need for better, process of recovering original image from the blurred image is
innovative and faster solutions still exits. The rationale behind known as Deconvolution process.
reviewing single blind image deblurring techniques is to
identify the research gap in the literature and contribute by
proposing a solution in the near future.

III. BLUR DEFINITION

Blur is a historical problem for which solutions have been
studied over a number of decades. Its root cause is the
inability of an imaging system to focus on objects on an image
plane at the same time, or relative motion between a camera
lens and an object during exposure. The uncontrolled and
varying conditions of forensic imagery can result in both of
these causes being realized, and thus it is extremely common Fig 3. Types of Image Deblurring Techniques
for both images and video to require some form of de-blurring
within the sequence of enhancement operators applied to Non blind Deconvolution is the process of estimating the
recover high-frequency information, and thus detail [8] latent image from a pair of latent & blurred image along with
[9][10][11]. the knowledge of PSF such as motion PSF or Gaussian PSF,
commonly used techniques are deconvolution using Lucy
Richardson, Regularized Filter & Weiner Filter [20]. Blind
Deconvolution is a technique that recovers the latent image
from a single or a set of blurred images where the point spread
function (PSF) type is unknown and hence involves estimation
of the PSF and restoration of the latent image as the
deconvolution is performed [21]. The process of image
degradation is described in Fig 4.
Figure 2a(left)-Blurred Images acquired from a CCTV surveillance camera
Figure 2b(right) - Blurred Images acquired from a video recorded on a
mobile phone [7].

The blurred image can be expressed as: G = H * F + N

where, F is the original latent image, H is the point spread Fig 4. Image Degradation Model [22]
function (PSF); it controls the degree of distortion in the
image, N is the additional noise that gets picked up during Analysis of various non-blind and blind deblurring techniques
acquisition of the image and is responsible for corruption of done by Yadav, Sudha et al [4], Devi, M et al [13] and
the latent image, (*) is the process of convolution between Muthana, R et al [14] are tabled in Table 1 using the metric of
latent image and PSF while G is the degraded blurred image Pixel Signal to Noise Ratio (PSNR) as metric. The following
[12]. Table 1 gives a comparative analysis of the different
A. Types of Blur: deblurring techniques mentioned above using the metric of
Following are the different types of blurs encountered in Pixel Signal to Noise Ratio (PSNR). The result of the analysis
the literature of image processing: gaussian blur, average blur, is visualized below in Fig 5. The analysis indicates that
motion blur, box blur, out of focus blur, atmospheric blur techniques in the category of non-blind image deblurring have
[13][14][15]. Restoration tasks in image forensics usually low PSNR values while techniques in the category of blind
identify motion blur and out-of-focus blur as the main source image deblurring have high PSNR values, namely ADSD-AR
of image blurring. Motion blur occurs due to the movement of gives highest PSNR value, followed by Neural Networks.
objects during the image acquisition process [16] or the
movement of cameras at the time of the scene capturing [17].
Lens defocusing is the major reason for out-of-focus blur [18].
B. Deblurring - Definition, Types, Techniques

ISSN: 2347-8578 www.ijcstjournal.org Page 130

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023

Table 1. Analysis of various Non-blind and Blind deblurring techniques done by

Yadav, Sudha et al [4], Devi, M et al [13] and Muthana, R et al [14].
Yadav, Sudha et al [4] Devi, M et al [13] Muthana, R et al [14]
Method Types of Perfor- PSNR Types of Perfor- PSNR Types of Perfor- PSNR
blur mance Ratio blur mance Ratio blur mance Ratio

Blind Gaussian Gaussian Efficient Gaussian &

Efficient 26.78 26.77 Efficient 26.78
deconvo-lution blur blur and good Motion
Lucy Gaussian Gaussian Gaussian
Efficient 21.06 Efficient 21.05 Efficient 21.06
Richardson blur blur blur

Worst
Gaussian Worst Gaussian Gaussian Worst
Wiener filter 17.05 performan 17.07 17.05
blur performance blur blur performance
ce

Regularised Gaussian Gaussian Gaussian

Efficient 20.1 Efficient 20.12 Efficient 20.1
filter blur blur blur
Handling
Gaussian Efficient 21.9 Gaussian Efficient 21.92 -- -- --
outliers
Very Very Very
ASDS-AR Gaussian 31.2 Gaussian 31.21 Gaussian 31.2
Efficient Efficient Efficient
Gaussian
Neural Gaussian Very Very Gaussian Very
30.11 Out-of- 30.1 30.11
Networks Out-of-focus Efficient Efficient Out-of-focus Efficient
focus

Hyper-
Hyper- Hyper-
Hyper-spectral spectra
spectra Efficient 22.34 Efficient 22.35 spectra Efficient 22.34
(PCA) l image
l image blur l image blur
blur

Motion density Motion Efficient 24.31 Motion Efficient 24.3 Motion Efficient 24.3

In general, blind deblurring techniques give better results than quality of the restored image. Hence, the PSNR values
non-blind deblurring techniques. The PSNR is a metric which depicted in Figure 5 indicates further scope for improvement
computes the peak signal to noise ratio between restored in devising better methods for blind image deblurring. The
image and the original image, it is measured in decibels. above analysis justifies the study of the different techniques
for blind image deblurring for identifying gaps in literature so
that better techniques can be created in the future.

Fig 5a. Charts summarizing the analysis of various Fig 5b. Charts summarizing the analysis of various
non-blind deblurring techniques blind deblurring techniques

PSNR serves as an image quality metric used for quality

evaluation of the restored image. Higher the PSNR, better the

ISSN: 2347-8578 www.ijcstjournal.org Page 131

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
IV. RELATED WORK Takahashi et. al. (2023) [23], consider blind deconvolution in
the Fourier domain. They express the deconvolution problem
About 69 research papers published in reputed journals
in the form of quartic term which does not have a Lipschitz
surveyed for understanding the recent literature on blind
continuous gradient. As the convergence analysis of existing
image deblurring methods have been categorized as shown in
first order methods depends on the existence of Lipschitz
the Fig 6.
continuous, traditional first order methods namely the fast

Fig 6. Categorization of papers surveyed on blind image deblurring methods.

iterative shrinkage thresholding algorithm (FISTA) are not

A. Papers on Maximum a posteriori (MAP) feasible. The authors propose to use Bregman-based proximal
Eqtedaei et al. (2022) [29] propose a multiscale blind gradient algorithms with extrapolation (BPDCAe) where they
image deblurring method based on maximum a posterior demonstrate to have found an appropriate difference of
(MAP) framework focusing on method that has less convex (DC) decomposition and kernel generating distance.
computational complexity and needs less execution time. With the motion blur process, Liu et al. (2022) [24] make use
K-means clustering is applied on the information about of the property of the patch wise second order gradient (PSG)
dominant edges of the blurred images are segmented in a of an image which happen to be decreasing in terms of its
coarse to fine scale. Mao et al. (2018) [84] propose a values. The authors propose a model with the PSG to handle
lightweight MAP model for blind image deblurring with the time-consuming trouble of the current prior- based blind
necessary constraints on latent image and blur kernel. The image deblurring methods. Initially, sub-sampled of the
authors use an adaptable sparse structure management for blurred image at multiple scales is performed then blur kernel
MAP inference using an imprecise projected gradient estimation and intermediate result are alternately iterated
approach. using PSG to get the final blur kernel. Finally, a suitable non-
B. Papers on priors blind approach is adopted to generate the final result. An

ISSN: 2347-8578 www.ijcstjournal.org Page 132

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
image prior driven by the dual principles of discriminativeness Selection (RBPS) model to overcome the presence of ringing,
(DPD) is more crucial and necessary to blind deblurring, staircase artefacts & loss of information leading to low quality
according to Shao et al. (2023) [37]. With the notion of DPD, in the deblurred images, which is frequently observed in most
a more concise and intuitive image model reweighted of the state-of-the-art image deblurring methods. The authors
Tikohonov regularization is presented and studied. propose their model is aimed to result in high quality
Particularly, the RTR model is transformed to a discriminative deblurred images without the formation of artefacts by using
filter and has been harnessed in a plug-and-play framework two algorithms. The First algorithm divides the image into
for blur kernel estimation. Liu et al. (2022) [42] propose a different regions based on the presence of blur so that an
fast-blind image deblurring algorithm which uses the salience appropriate prior can be applied to each region. Then different
map along with gradient cepstrum. The authors take priors for each image region are applied by using a hybrid
advantage of the recently discovered phenomenon that the regularization algorithm.
extreme values of the salience map of the clear image are Zhang et al. (2022) [48] propose deblurring algorithm for
sparser than the blurred one, and as a result, they use a L0- a remote sensing images based on Local Binary Pattern (LBP)
regularized salience map prior with the half-quadratic splitting prior that prevents over-sharpening by classifying all pixels
algorithm for optimisation and a gradient cepstrum to prevent and processing them in different ways. The authors map the
too many iterations in each scale and to initialise the blur pixels containing the crucial texture features of the images
kernel. The simplified extreme channel prior (SECP) and that do not change in the presence of blur using local binary
gradient cepstrum used in Liu et al.'s (2022) [56] fast blind patterns (LBP). The fast-iterative shrinkage-thresholding
image deblurring algorithm are modelled after and inspired by algorithm (FISTA), the half-quadratic splitting method, and
the clear image's simplified bright channel prior (SBCP) and the projected alternating minimization (PAM) algorithm are
simplified dark channel prior (SDCP). By imposing the L0 used to create the mapping matrix.
norm constraint on the terms involving SECP and Li et al. [59] (2021) proposes an edge extraction module
incorporating them into the conventional deblurring based on L0 sparse representation for blind motion deblurring
framework, the authors optimise the model. An gradient which preserves the edges of images, which are embedded in
cepstrum is used to determine the initial kernel size and to a multi-scale recurrent network (SRN). Edge enhancement is
prevent excessive iterations in each scale. A gradient-based performed when the current scale transmits information to the
discriminative prior is suggested by Shao et al. (2020) [65] for next scale. The authors propose an dual-attention mechanism
precise and reliable blur kernel estimation. The authors have into the encoder-decoder structure which exploits the
used the concept of dual principles of discriminativeness correlation among pixels and the correlation among channels.
(DPD) for the blind deblurring problem. vii. Latent Structure prior: The single-image blind deblurring
Yu et al. (2022) [28] propose a Fractional Order Dark method proposed by Bai et al. (2020) [69] uses MultiScale
Channel Prior (FODCP) for image deblurring based on the Latent Structures (MSLS) to restore sharp images in the
fact that a fractional order calculation can inhibit the noise and coarse scales before using an enhancement process in the fine
preserve the texture information present in the image. The scale to produce the final deblurred image. By alternately
proposed method can produce an efficient dark channel of a carrying out a sharp image reconstruction with a quick local
degraded image containing blur/noise and can approximate self-example matching and an accelerated kernel estimation
the blur kernel by suppressing the noise in the image. Full- with error correction, the computational complexity is reduced
reference & No-Reference metrics are used for assessment of at each scale.
image quality. To deal with the task of restoring remote sensing images,
Li et al. (2023) [31] propose a joint sparsity prior and Zhu et al. (2022) [40] propose an entirely new prior called the
multi-scale fusion model for blind text image deblurring, overlapped patches non-linear (OPNL) prior. It is derived
which contains the sparsity priors for estimating latent clean from the ratio of extreme pixels affected by blurring in
image gradient, blur kernel intensity and high frequency patches where the projected alternating minimization (PAM)
wavelet coefficients of the clear latent image. On the basis of algorithm is used along with other techniques like the fast
Laplacian weight and saliency weight, the multi-scale fusion iterative shrinkage-thresholding algorithm (FISTA), fast
technique reduces the influence of the brightness feature of Fourier transform (FFT), and others. Ljubenović et al. (2017)
the restored blur kernel. In order to solve the BID problem, [75] propose a method called a Gaussian mixture model
Yang et al. (2022) [72] suggest building a novel sparse (GMM) that uses patch-based priors to learn a class adapted
channel prior derived from the ratio of the dark channel prior prior by training on a dataset of clean images of that class.
(DCP) to the bright channel prior (BCP). The authors use the The authors use a plug-and play approach with a denoiser
auxiliary variable technique to incorporate the scant prior data based on Gaussian mixture model (GMM) and use a weaker
into the iterative restoration process. A blind image deblurring prior on the blurring filter.
technique based on sparse representation that makes use of the Bai et al. (2019) [77] propose a graph-based image prior
sparsity property of natural images is proposed by Zhang et al. RGTV that promotes a bi-modal weight distribution to
(2011) [78]. It assumes that an over-complete dictionary can reconstruct a skeleton patch from a blurry observation, so that
be sparsely represent the patches from the natural images. a suitable blur kernel can be simply derived thereafter. The
S, Pooja et al. (2022) [35] propose a Region Based Prior authors design an efficient algorithm to solve the non-convex

ISSN: 2347-8578 www.ijcstjournal.org Page 133

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
nondifferentiable optimization problem alternately and adaptively changes the parameter values during the iterative
suggest the use of accelerated graph spectral filters as a process and helps to better maintain the details in the
speedup method for blind Gaussian deblurring. recovered image. Ge et al.'s [55] proposal for a blind image
x. Tri-segment intensity prior: Zhang et al. (2020) [79] deblurring algorithm makes use of the image surface's
propose a blind image deblurring method based on the tri- Gaussian curvature (GC). By minimising the combined
segment intensity prior to estimate a blur kernel from a single regularization, the authors show that GC regularization, when
motion blurred image. In order to estimate the blur kernel, the combined with L0-norm of image gradients, is effective for
sharp image and the blur kernel are alternately estimated. The blind image deblurring while maintaining sharp edges and
local minimal solution is significantly avoided in a coarse-to- removing unwanted structures and noises in intermediate
fine manner using an image pyramid. latent images. The half-quadratic splitting technique and the
Lui et al. (2020) [82] uses the Local Maximum Difference Gaussian curvature filter (GCF) are used by the authors to
Prior (LMD) for blind image deblurring. The authors adopt optimise the solution. Duan et al. (2021) [57] attempt to solve
the L1 norm constraint to the LMD involved term, compute blur induced by atmospheric turbulence by proposing
LMD using a linear operator, and then use a half-quadratic suppressed projected alternating minimization (SPAM) a non-
splitting strategy to optimise the method. blind image deblurring algorithm that uses maximizing L1
Chen et al. (2020) [83] present a blind image deblurring regularization (ML1) an convex image prior. The authors'
method using on the non-local self-similarity, which can deblurring algorithm employs a soft suppression technique to
simultaneously capture the intrinsic structure correlation and get rid of the artefacts. Hsieh et al. (2021) [64] propose an
spatial sparsity of an image. The authors model the structure effective variational model for blind image deblurring where
information of non-local similar patches using the hyper- the authors implement a zero patch minimum constraint for
Laplace prior and insert it into the low rank model. restoring the latent clear image. Important features of the
Chen et al. in (2019) [85] propose a blind deblurring latent image are restored by assigning the patch minimum
method based on Local Maximum Gradient (LMG) prior. The information from the blurred image back to itself. An adaptive
authors use a linear operator to compute the Local Maximum regularizer is used to perform the final image restoration of
Gradient, together with an effective optimization scheme. blurred images having significantly better edge preserving
Cheng et al. (2020) [70] propose a single-image blind property than the total variation regularizer for image
deblurring framework called Hybrid Deep Priors Model restoration.
(HDPM) for kernel estimation and deblurring. The authors D. Papers on Parametric PSF estimation
propose using an implicit deep prior and hand-crafted explicit Nasonov et al. (2022) [27] focus on the estimation of blur
prior as regularizations into the MAP inference process to parameters. The authors use a patch-based linear
extract the detailed texture and sharp structures of latent approximation to motion blur for estimating the direction of
image. Li et al. (2018) [81] propose using a data-driven linear blur and a CNN model for estimating the parameters of
discriminative prior for blind image deblurring. The image a linear blur kernel. Li et al. (2023) [31] propose a joint
prior is suggested by the authors as a deep convolutional sparsity prior and multi-scale fusion model for blind text
neural network (CNN) binary classifier. By incorporating the image deblurring. The model employs wavelet coefficients
learned prior into the maximum a posterior (MAP) framework, with high frequency in the latent clean image, intensity of the
blind image deblurring is accomplished. The authors use the blur kernel, and sparsity priors to restore the gradient of the
half-quadratic splitting method and gradient decent algorithm latent clean image. On the basis of Laplacian weight and
to optimise the deblurring method because the CNN is non- saliency weight, the multi-scale fusion technique reduces the
linear. influence of the brightness feature of the restored blur kernel.
C. Papers on Regularization Parvaz (2021) [47] presents an algorithm to estimate the point
Dong et al. (2022) [43] propose a blind deblurring method spread function based on framelet transform and l0−αl1 norm
specifically for recovering clear Optical coherence regularisation for coarse-to-fine iteration. The algorithm uses
tomography (OCT) images from blurred images without a the framelet transform to improves the restored kernel which
known PSF using spatially adaptive approach. The authors is decomposed to the different frequencies and also employs a
first derive a depth-dependent PSF from the Gaussian beam fraction gradient operator on images with various PSF sizes
model then express the blind deconvolution as a regularized rather than a regular gradient operator.
energy minimization problem using the least squares method E. Papers on Kernel Estimation
followed by an alternating optimization method. To speed up Bai et al. (2022) [46] propose a blind single image
the execution the authors have used an accelerated alternating deblurring called the dual-channel block-cluster deblur
optimization method based on the convolution theorem and algorithm (DCBCD) for specifically restoring space-variant
Fourier transform. Zhang et al. (2022) [45] propose the blurred images. The model uses the characteristics offered by
Gaussian-scale mixture expert field (GSM-FOE) model as a the traditional multi image restoration image algorithm and a
regularised image recovery model which can better fit the combines it with those of the single image blind
higher-order prior of natural images and accurately portray the deconvolution restoration algorithm. The authors outline the
global prior knowledge of natural images. The authors suggest relative side prior information (RSPI) and kernel estimation
using IRLS, or iterative reweighted least squares which using non-dimensional weight-balanced Gaussian measure

ISSN: 2347-8578 www.ijcstjournal.org Page 134

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
(WG-NGM) and propose a clustering-based algorithm is lower bound and then use inference networks to approximate
proposed, to restore the space-variant blurred images. Sanghvi the corresponding posteriors of the clean image and blur
et al. (2022) [50] propose an unsupervised & iterative blind kernel in an amortized manner efficiently from an observed
deconvolution method for deblurring of low light images. The blurry image. Bao et al. (2021) [66] propose a method for
authors use a deep learning based non-blind Poisson solver to blind image deblurring and blur kernel estimation. The
estimate the latent image. The updation of the blur kernel is authors use a convolutional neural network (CNN) with a
carried out in an iterative manner where unsupervised loss model-based optimization method to solve the problem of
computed by propagating the gradient of the nonblind solver blind image deblurring. A variable splitting technique called
is backpropogated. Cao et al. (2022) [51] propose three-stage half quadratic splitting method (HQS) is used to combine the
restoration framework for deblurring a single blind image network model trained by discriminative learning approach
when there is noise in the form of a line pattern. Before with model-based optimization method to restore blind images.
estimating an intermediate blur kernel, the authors first Using a hierarchical architecture that is improved from U-Net,
eliminate the detrimental effects of the stripe noise in the Chi et al. (2022) [71] propose the single-image blind
unidirectional gradient domain, then the missing details of the deblurring framework Selective Residual M-Net (SRMNet).
kernel are rectified using a learning based kernel refinement The multi-scale semantic information is enhanced by the
network and finally, deblurring and noise removal are authors using a selective kernel with residual block on the
performed using a non-blind deblurring model based on a low hierarchical structure known as M-Net. In order to recover
rank decomposition. Zhu et al. (2020) [67] suggest a local images from hazy satellite images, Biyouki et al. (2021) [73]
rank-based single-image blind deblurring technique. The propose a kernel mixture structure using the Gaussian kernel
authors implement adaptive threshold segmentation on a as the base kernel, where multiple Gaussian kernels are
traditional local rank transform to create a blind image combined structurally and improved in terms of scales and
deblurring model then a half-quadratic splitting method is centres. Nan et al. (2020) [86] propose a framework of
used in alternating iterations for the estimation of the blur variational expected maximization (VEM) to solve noise-blind
kernel as well as the deblurred image. To create the desired image deblurring. The authors propose a deep-learning-based
latent image, a linear combination of the hyper-Laplacian computational scheme using a CNN to learn image prior and
model and the total-variation-l2 model is used. Tian et al. uses a multi-layer perception network (MLP) to quantify prior
(2021) [87] propose a blind image deblurring method based uncertainty. Li et al. (2021) [88] propose Multiple Auxiliary
on dual attention deep image prior (DADIP) network and 2- Networks (MANet) for single blind image deblurring method
dimensional (2D) blur kernel estimation with convolutional which assists norm L1-loss function and enhances the quality
neural network (CNN). The authors focus on eliminating the of the deblurring image. The main branch of MANet is an
ringing artifacts and wrinkles that appear after deburring. The encoder-decoder structure made up of residual blocks, and the
dual attention mechanism is combined with squeeze and three auxiliary branches are the edge prediction branch, the
excitation network (SENet), which greatly improves the multi-scale refinement branch, and the perceptual loss branch.
restoration effect of image details while the 2D blur kernel Zhuang et al. (2022) [90] identify issues of SOTA single-
estimation approach via CNN suppresses the ringing artifacts instance BID methods and propose a revamped SelfDeblur
of the image. with a number of crucial modifications to resolve these issues.
F. Deep Learning based papers The authors attempt to address the issue with their model by
Mao et al. (2023) [33] propose a deep idempotent pointing out that several state-of-the-art (SOTA) single-
network for efficient single image blind deblurring that uses instance methods are unstable when the kernel size is
the idempotent constraint to the deep deblurring network, overspecified and/or the noise level is high. An et al. (2020)
which improves the non-uniform deblurring performance and [91] propose a blind single image deblurring method Blur
achieves stable results w.r.t. re-deblurring multiple times. The Invariant Kernel-Adaptive Network (BIKAnet) that uses a
authors have designed a deblurring network through kernel estimation network that produces adaptive blur kernels
progressive residual deblurring with recurrent structure. In based on the analysis of the blurred image and a deblurring
order to improve document-image quality by reducing the network that restores sharp images using the estimated blur
impact of distortions present in them, Mohsenzadegan et al. kernel.
(2022) [34] propose a new convolutional neural network To learn the deblurring from multi-modal medical images,
(CNN) architecture. The authors use two modules, first for Sharif et al. (2023) [38] propose an end-to-end scale-recurrent
deblurring and the second for jointly denoising and contrast deep network combining residual dense block and spatial
enhancement. A deep variational Bayesian framework asymmetric attention (RD-SAM). The residual-dense learning
(VBDeblurNet) is proposed by Zhao et al. (2022) [44] for the strives to extract salient features, and the spatial-asymmetric
blind deblurring of images, in which deep neural networks can block refines the extracted features with local-global attention.
be used to jointly estimate the posterior of the latent clean A framework for image deblurring that focuses on the
image and blur kernel. The authors construct the variational reconstruction of high-frequency information is proposed by
distributions parameterized by DNNs to approximate the true Liu et al. (2022) [53]. The proposed model implements
posteriors of the latent clear image as well as the blur kernel deblurring processes with high-frequency guidance at various
and learn the network parameters by optimizing the evidence scales simultaneously using a high-frequency reconstruction

ISSN: 2347-8578 www.ijcstjournal.org Page 135

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
subnetwork (HFRSN), a high-frequency reconstruction approaches. The proposed method uses wavelet thresholding,
subnetwork (HFRSN), and a multi-scale grid subnetwork multi-scale forward transform, denoising of the detail
(MSGSN). Using a high-frequency information aggregation coefficients and an inverse transform which is non-linear
(HFAG) module and a high-frequency information attention transform based on the lifting scheme that inherits the
(HFAT) module to direct the feature extraction stage and the sparsifying ability, perfect reconstruction property as well as
feature reconstruction stage in each scale of MSGSN, the multi-scale property of the wavelet transform.
respectively, the authors also suggest restoring sharpening in Xu et al. (2022) [32] propose a method called Cycle GAN
images. that can learn the mapping between the blurred Millimeter-
Li et al. (2022) [36] propose a dual branch discriminator, Wave (MMW) images and the focused ones without using
which can discriminate image authenticity and distinguish paired training samples. MMW imaging is used to scan human
image sparsity simultaneously using a dual-branch GAN bodies for concealed weapon detection where the slightest
which employs various image sparse priors. A new end-to-end movement introduces blur. The authors investigation of
generative adversarial network (SDD-GAN) for single-image blurring in MMW image due to the shaking of the human
motion deblurring of remote sensing images is proposed by body is the first attemp ever to address this problem.
Zhang et al. in 2022 [39]. The authors propose a Feature Nasr et al. (2022) [41] proposes to use game theory to
Alignment Module (FAFM) and Feature Importance Selection solve the problem of blind deconvolution using the the Kalai
Module (FISM) to successfully recover the texture and details Smorodinsky solution solved geometrically using the normal
of the blurred remote sensing images. Tomosada et al. (2021) boundary intersection (NBI) method.
[62] proposes a high quality image deblurring method for A transform-based method for estimating the linear
large motion blurs called DeblurDCTGAN that preserves motion blur parameters is proposed by Jaravan et al. (2021)
texture and suppresses ringing artifacts in the restored image [54] based on the extraction of features from the provided
without multi-scale or multi-patch architecture using DCT single blurred image. To determine the direction of motion
based loss. The algorithm uses discrete cosine transform blur, the Radon transform of the blurred image's spectrum is
(DCT) and requires less computational complexity. used. The length of the motion blur is calculated using the
Frequency-domain Contrastive Loss Constrained Lightweight Noise-Immune Discrete Cosine Transform-based (NIDCT)
CycleGAN, or simply FCL-GAN, is a lightweight and real- blur metric.
time unsupervised BID baseline that Zhao (2021) [87] et al.
propose. FCL-GAN is faster and lighter and has no V. OBSERVATIONS
restrictions on the image domain or resolution. To make the
The following observations have been made based on the
model smaller and run more quickly, the authors use the
extensive literature survey done on papers published in recent
lightweight domain conversion unit (LDCU) and parameter-
years presented in section 4 of this paper.
free frequency-domain contrastive unit (PFCU).
A. Categories of techniques used
G. Papers from other categories
Over the years, image deblurring algorithms have made
Lee et. al. (2022) [26] proposes a framework for image
significant breakthroughs leading to the development of a
restoration improving the quality of the PET image by
number of methods and categories of methods. Classification
combining blind deconvolution-based deblurring and noise
of image deblurring techniques are summarized in Table 2.
reduction using a non-local means (NLM) strategy. Almeida
et al. (2011) [80] suggest a blind image deblurring method to Table 2: Classification of Image Deblurring Techniques
overcome the drawback of BID methods, which is that
Classification Categories of techniques
iterations must be manually stopped. BID methods do not
Criteria
require prior knowledge of the blurring filter. On the basis of
Based on knowledge Blind deblurring, non-blind
this reasoning, the authors suggest that if the estimation blur
of the blur kernel deblurring
filter is effective, the residual will be spectrally white, but if it
Based on image input Single frame, multi frame
is not, the deblurred image will contain structured artefacts
that are not white. Based on degradation Motion blur, gaussian blur,
type defocus blur, turbulence blur, etc.
Hu et al. (2022) [30] in order to deblur blind images, Based on method of Optimization-based deblurring
suggest a powerful mixed-order salient edge selection. The deblurring algorithm method and deep learning
authors use zero-order information to retrieve refined deblurring method
structures from the blurred image, then use the second-order Based on consistency Space-changing blur and space-
information to retrieve the strong structure of salient edges of degradation invariant blur
and provide and accurate blurred kernel estimation. Then a
combination of three orders is used for blind image deblurring
which proves to be a robust salient edge method. B. Image Quality Metrics used
Huang et al. (2022) [52] propose a blind deblurring Considering the increasing applications of image processing,
method called wavelet inspired invertible network (WINNet). enhancement & restoration there is a lot of research being
This method integrates the wavelet and learning based done in the computational Image Quality Assessment (IQA)
methods. Researchers use different Image Quality Metrics

ISSN: 2347-8578 www.ijcstjournal.org Page 136

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
(IQM) to assess the quality of output images produced by their blur and hence introduce blur/noise to a clean image and
model. The many image quality metrics can be categorized evaluate the performance of their algorithms on the distorted
into two broad categories – subjective & objective are image. Benchmark datasets specifically created for evaluating
summarized in Figure 7. In subjective IQM’s, the given the performance of deblurring algorithms are summarized in
image(s) is evaluated by human observers through subjective Table 3.
evaluation by assigning scores to them and hence quantifying
the visual quality of an image. These methods are time D. State-of-the-art & benchmark deblurring algorithms
consuming hence inconvenient for real life implementation Fergus et al. [92] (2006) propose a technique for removing
and cannot be integrated with automated tools. In objective the effects of unknown camera shake from an image. The
IQM’s the main purpose is to use mathematical models to authors use natural image priors and advanced statistical
automatically and accurately evaluate the quality of an image. techniques focused on kernel estimation. The results however
Objective IQMs can be grouped according to the concept of indicate presence of ringing artifacts. Shan et al. [110] (2008)
reference image or use of Human Visual System (HSV) propose a technique for single image motion deblurring by
models. Based on the concept of a reference image objective first exploring the major causes of visual artifacts and then
IQMs can be subcategorized into three classes: Fully- apply advanced iterative optimization that, until convergence,
Reference (FR), Reduced-Reference (RR) and No-Reference alternates between blur kernel enhancement and image
(NR). FR IQM’s evaluate the quality of a pair of images restoration. The proposed method is a probabilistic model that
containing the ground-truth and the distorted images. unifies blind and non-blind deconvolutions into a single MAP
Some of the FR IQMs are Mean Square Error (MSE), Mean formulation. Cho & Lee [93] (2009) propose a fast deblurring
absolute Difference (MD), Pearson Correlation Coefficient method which works with image derivatives rather than pixels
(PCC), Spearman Rank Correlation Coefficient (SRCC), Peak to accelerate the iterative deblurring process, both latent
signal-to-noise-ratio (PSNR). Reduced-reference IQMs assess image estimation and kernel estimation. Xu & Jia [96] (2010)
the quality of images which have the distorted images and propose a motion deblurring method which adaptively select
some additional information about the original ground truth useful edges for kernel estimation and uses the iterative
image. NR IQM’s assess the quality of images which only support detection (ISD) based kernel refinement with adaptive
have the distorted image without any other prior information regularization to improves the result quality. Whyte et al. [102]
of how the free-distortions image should look like. Some of (2010) propose a deblurring method which is a parametrized
the NR IQM’s are Blind / Reference Less Image Spatial geometrically consistent model of non-uniform image blur.
Krishnan et al. [98] (2011) propose a fast blind deblurring

Figure 7: Classification of Image Quality Metrics (IQM’s)

Quality Evaluator (BRISQUE), Perception based Image

Quality Evaluator (PIQE), Natural Image Quality Evaluator method which uses a new type of non-convex image
(NIQE), Blind Image Integrity Notator using DCT Statistics regularization which favours sharp images over blurry ones
(BLIINDS), Distortion Identification Based on Image and provides the sharpest image at the lowest cost. The ratio
Identification and Integrity Test (DIIVINE). The constraints of the l1 norm to the l2 norm on an image's high frequencies is
on size of the paper restrict the detailed description of these the suggested regularisation function. Levin et al. [100] (2011)
metrics as well as numerous other IQM’s in the literatutre. propose an algorithm which optimizes the MAP algorithm and
C. Benchmark Datasets performs alternating minimization to solve for the original
Several datasets have been used by researchers for sharp image and the true blur kernel. Hirsh et al. [109] (2011)
evaluating the performance of deblurring algorithms. Most present a blind deblurring algorithm on single image for
researchers have used existing datasets which are specific to a removing motion blur of non-uniform nature induced by
certain domain like medical imaging, remote sensing, facial camera shake and uses Efficient Filter Flow (EFF) &
images, textual images, etc which originally do not contain Projective Motion Path Blur models (PMPB models) to

ISSN: 2347-8578 www.ijcstjournal.org Page 137

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
achieve the results. A generalised and sound mathematical L0 blurry & sharp stereo image pairs along with their
sparse expression is used in Xu's (2013) motion deblurring bidirectional disparities. Kupyn et al. [99] (2019) propose
method (p. 78). Xhang [95] (2013) propose a multi-image DeblurGAN-v2, an improvement over DeblurGAN [112] for
blind deblurring method which uses a penalty function that single image motion deblurring. Gao et al. [100] (2019)
combines the latent sharp image, blur kernels, and noise propose a method for dynamic scene deblurring using feature
variances. By applying a series of directional filters to the transformation modules with nested skip connection structures.
input image, Zhong [97] (2013) propose a new method for The authors analyze the principles of effectively using
dealing with noise in blind image deblurring. This method parameters in deblurring CNNs and propose a parameter
uses the inverse Radon transform to reconstruct the blur selective sharing scheme in contrast to parameter
kernel and estimates an accurate Radon transform of the blur independence and sharing schemes. Ye et al. [102] (2019)
kernel from each filtered image. Sun et al. [104] (2013) propose a method for blind image deblurring called scale-
propose a kernel estimation method which uses an edge-based iterative upscaling network (SIUN). It uses weights sharing
approach using patch priors on edges of the latent image. The across iterations. Wen et al. [92] (2020) suggests patch-wise
authors also propose a new synthetic test set of 640 high- minimal pixels (PMP), a sparsity prior of local minimal pixels
resolution natural images of diverse scenes. Michaeli & Irani having the property that a clear image with more sparsity will
[99] (2014) propose a blind-deblurring method based on the have a larger PMP of than that of blurred image and so can be
internal patch recurrence property within a single natural used to differentiate between clear and blurred images. In Qi
image used as a cue for estimation of the blur kernel. Pan et al. et al.'s [101] (2020) proposal, we use a dense feature fusion
[101] (2014) propose exemplar-based method for deblurring block made up of channel attention and pixel attention
face images which exploits the structural information. kernel modules to propose a blind image deblurring method. The
estimation is done using face structure and reliable edges from authors employ scale attention modules and densely
exemplars. Total variation is used by Perrone et al. [103] to connected multiple dense feature fusion blocks to acquire
analyse the application of sparse gradient priors. The authors high-order feature representation.
confirm that the problem formulation of total variation blind E. Publication Statistics
deconvolution as a maximum a priori in both sharp image and This section presents the bifurcation of surveyed papers
blur is prone to local minima. The dark channel of the according to publisher & year of publication and is presented
recovered image is minimised using a L0-regularization term in Figure 8.
in Pan et al. [105] (2016)'s proposal for a straightforward and
efficient blind image deblurring method based on the dark
channel prior. For text image deblurring, Pan et al. [106] V. CONCLUSIONS
(2016) suggest an efficient L0-regularized prior based on
This paper is based on a survey of 69 research papers in last 5
intensity and gradient. The authors construct a 120-image
years. The survey is constrained by the availability of quality
dataset containing 15 ground truth document images and 8
research papers. Categorization of the different deblurring
kernels. Yan et al. [107] (2016) propose a joint prior by
algorithms according to possible criteria is presented in Table
combining Bright and Dark Channel Prior called extreme
2, categorization of frequent (not all) Image Quality Metrics
channels prior (ECP). Nah et al.'s [111] (2017) non-uniform
(IQM’s) is presented in Figure 7. Frequently used benchmark
blind deblurring method directly restores latent images
datasets for evaluating the performance of deblurring
without relying on any constrained blur kernel model, and it
algorithms is presented in Table 3. A short review of state-of-
employs a multi-scale convolutional neural network. The
the-art & benchmark deblurring algorithms used for
proposed model follows a coarse-to-fine approach and is
comparative quantitative analysis is presented in section 5D.
trained in multi-scale space. The authors construct a dataset is
composed of 3214 pairs of blurry and sharp images at
1280x720 resolution. Kupyn et al. [112] (2016) propose
DeblurGAN for blind motion deblurring based on a
conditional GAN and optimized using a multi-component loss
function. Tao et al. [113] (2018) present a coarse-to-fine blind
image deblurring technique called Scale-recurrent Network
(SRN-DeblurNet). Chen et al. [85] in (2019) propose a blind
deblurring method based on Local Maximum Gradient (LMG)
prior. The authors use a linear operator to compute the Local
Maximum Gradient, together with an effective optimization
scheme. Zhou et al. [114] (2019) propose a stereo image
deblurring method with Depth Awareness and View
Aggregation network (DAVANet). The authors combine the
estimation of bidirectional disparities with deblurring using a
fusion network. The authors have created and shared a dataset
for the purpose of stereo deblurring which contains20,637

ISSN: 2347-8578 www.ijcstjournal.org Page 138

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023

Figure 8: Bifurcation of surveyed papers according to publisher & year of publication

Deblurring.” 2017 IEEE International Conference on
Computer Vision (ICCV) (2017): 231-240.
VII. REFERENCES [4] Yadav, Sudha et al. “Evaluation of Image Deblurring
Techniques.” International Journal of Computer
[1] Bestagini, Paolo et al. “An overview on video
Applications 139
forensics.” 2012 Proceedings of the 20th European
Signal Processing Conference (EUSIPCO) (2012):
1229-1233.
[2] Jain, Sindhu and Mr. Sudhir Goswami. “A
comparative study of various image restoration
(2016): 32-36.
techniques with different types of blur.” (2015): Vol.3
[5] Ali, Usman and Muhammad Tariq Mahmood.
Issue 11, Pg.: 54-60.
“Analysis of Blur Measure Operators for Single
[3] Wieschollek, Patrick et al. “Learning Blind Motion

ISSN: 2347-8578 www.ijcstjournal.org Page 139

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
Image Blur Segmentation.” Applied Sciences 8 Deblurring.” Traitement du Signal 37 (2020): 527-
(2018): 807. 539.
[6] Jain, Sindhu and Mr. Sudhir Goswami. “A [23] Takahashi, Shota et al. “Blind deconvolution with
Comparative Study of Various Image Restoration non-smooth regularization via Bregman proximal
Techniques with Different Types of Blur.” (2015). DCAs.” Signal Process. 202 (2022): 108734.
[7] Russ, J. C. (2001, May 25). Forensic Uses of Digital [24] Liu, Jing et al. “Blind deblurring with patch-wise
Imaging. https://doi.org/10.1604/ 9780849309038 second-order gradient prior.” Signal Process. Image
[8] Zjalic & Verden Forensics. “Blur.” (2020): Commun. 107 (2022): 116781.
https://foclar.com/news/blur. Retrieved March 3, [25] Liu, Jing et al. “Blind deblurring with fractional-order
2023, from https://foclar.com/news/blur calculus and local minimal pixel prior.” J. Vis.
[9] Ben-Ezra, Moshe and Shree K. Nayar. “Motion-based Commun. Image Represent. 89 (2022): 103645.
motion deblurring.” IEEE Transactions on Pattern [26] Lee, Min-Hee et al. “Image restoration algorithm
Analysis and Machine Intelligence 26 (2004): 689- incorporating methods to remove noise and blurring
698. from positron emission tomography imaging for
[10] Al-Ameen, Zohair et al. “Restoring Degraded Alzheimer's disease diagnosis.” Physica medica :
Astronomy Images using a Combination of Denoising PM : an international journal devoted to the
and Deblurring Techniques.” (2012). applications of physics to medicine and biology :
[11] Al-Ameen, Zohair et al. “Computer Forensics and official journal of the Italian Association of
Image Deblurring: An Inclusive Biomedical Physics 103 (2022): 181-189 .
Investigation.” International Journal of Modern [27] Nasonov, Andrey V. and Alexandra A. Nasonova.
Education and Computer Science 5 (2013): 42-48. “Linear Blur Parameters Estimation Using a
[12] Pan, Jin-shan et al. “Learning to Deblur Images with Convolutional Neural Network.” Pattern Recognition
Exemplars.” IEEE Transactions on Pattern Analysis and Image Analysis 32 (2022): 611-615.
and Machine Intelligence 41 (2018): 1412-1425. [28] Yu, Xiaoyuan et al. “A Single Image Deblurring
[13] Devi, M. Kalpana and R. Ashwini. “An Analysis on Approach Based on a Fractional Order Dark Channel
Implementation of Various Deblurring Techniques in Prior.” International Journal of Applied Mathematics
Image Processing.” (2016). and Computer Science 32 (2022): 441 - 454.
[14] Muthana, Rasha and Assad N. Alshareefi. [29] Eqtedaei, Amir and Alireza Ahmadyfard. “Coarse-to-
“Techniques in De-Blurring Image.” Journal of fine blind image deblurring based on K-means
Physics: Conference Series 1530 (2020): n. pag. clustering.” The Visual Computer (2023): n. pag.
[15] Sulong, Ghazali. “Toward a better understanding of [30] Hu, Dandan et al. “Image Deconvolution Using
deblurring digital images in computer forensics.” Mixed-Order Salient Edge Selection.” Circuits,
(2013). Systems, and Signal Processing (2023): n. pag.
[16] Potmesil, Michael and Indranil Chakravarty. [31] Li, Zhe et al. “Blind Text Image Deblurring
“Modeling motion blur in computer-generated Algorithm Based on Multi-Scale Fusion and Sparse
images.” Proceedings of the 10th annual conference Priors.” IEEE Access 11 (2023): 16042-16055.
on Computer graphics and interactive [32] Liu, Huteng et al. “Millimeter-Wave Image
techniques (1983): n. pag. Deblurring via Cycle-Consistent Adversarial
[17] Raskar, Ramesh et al. “Coded exposure photography: Network.” Electronics (2023): n. pag.
motion deblurring using fluttered shutter.” ACM [33] Mao, Yuxin et al. “Deep Idempotent Network for
Trans. Graph. 25 (2006): 795-804. Efficient Single Image Blind Deblurring.” IEEE
[18] Moghaddam, Mohsen Ebrahimi. “Out of focus blur Transactions on Circuits and Systems for Video
estimation using genetic algorithm.” 2008 15th Technology 33 (2022): 172-185.
International Conference on Systems, Signals and [34] Mohsenzadegan, Kabeh et al. “Deep Neural Network
Image Processing (2008): 417-420. Concept for a Blind Enhancement of Document-
[19] Kaur, Ravneet and Er. Navdeep Singh. “Image Images in the Presence of Multiple
Restoration - A Survey.” IOSR Journal of Computer Distortions.” Applied Sciences (2022): n. pag.
Engineering 16 (2014): 107-111. [35] S, Pooja et al. “Image region driven prior selection for
[20] Cho, Sunghyun et al. “Handling outliers in non-blind image deblurring.” Multimedia Tools and
image deconvolution.” 2011 International Conference Applications (2023): n. pag.
on Computer Vision (2011): 495-502. [36] Li, Jialuo et al. “Dual-Branch Discrimination Network
[21] Lam and Goodman. “Iterative statistical approach to Using Multiple Sparse Priors for Image
blind image deconvolution.” Journal of the Optical Deblurring.” Sensors (Basel, Switzerland) 22 (2022):
Society of America. A, Optics, image science, and n. pag.
vision 17 7 (2000): 1177-84 . [37] Shao, Wenze et al. “RTR-DPD: reweighted tikohonov
[22] Satish, Pooja et al. “A Comprehensive Review of regularization for blind deblurring via dual principles
Blind Deconvolution Techniques for Image of discriminativeness.” Multidimensional Systems and

ISSN: 2347-8578 www.ijcstjournal.org Page 140

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
Signal Processing (2023): 1-30. “Using a Blur Metric to Estimate Linear Motion Blur
[38] Sharif, S. M. A. et al. “MedDeblur: Medical Image Parameters.” Computational and Mathematical
Deblurring with Residual Dense Spatial-Asymmetric Methods in Medicine 2021 (2021): n. pag.
Attention.” Mathematics (2022): n. pag. [55] Ge, Xianyu et al. “Blind image deblurring with
[39] Zhang, Ziyu et al. “Blind Remote Sensing Image Gaussian curvature of the image surface.” Signal
Deblurring Based on Overlapped Patches’ Non-Linear Process. Image Commun. 100 (2021): 116531.
Prior.” Sensors (Basel, Switzerland) 22 (2022): n. pag. [56] Liu, Jing et al. “Gradient Cepstrum Combined with
[40] Zhu, Baoyu et al. “Blind Deblurring of Remote- Simplified Extreme Channel Prior for Blind
Sensing Single Images Based on Feature Deconvolution.” Circuits, Systems, and Signal
Alignment.” Sensors (Basel, Switzerland) 22 (2022): Processing 41 (2021): 1074-1099.
n. pag. [57] Duan, Lizhen et al. “Deblurring Turbulent Images via
[41] Nasr, Nora et al. “The Kalai Smorodinsky solution for Maximizing L1 Regularization.” Symmetry 13
blind deconvolution.” Computational and Applied (2021): 1414.
Mathematics 41 (2022): n. pag. [58] Fiori, Simone G. O.. “Improvement and Assessment
[42] Liu, Jing et al. “A fast blind image deblurring method of a Blind Image Deblurring Algorithm Based on
using salience map and gradient cepstrum.” The Independent Component Analysis.” Comput. 9
Visual Computer (2022): n. pag. (2021): 76.
[43] Dong, Wenxue et al. “Spatially adaptive blind [59] Li, Menghang et al. “Blind motion deblurring via L0
deconvolution methods for optical coherence sparse representation.” Comput. Graph. 97 (2021):
tomography.” Computers in biology and medicine 147 248-257.
(2022): 105650 . [60] Sun, Shuhan et al. “Blind Deblurring Based on
[44] Zhao, Qian, et al. “A Deep Variational Bayesian Sigmoid Function.” Sensors (Basel, Switzerland) 21
Framework for Blind Image Deblurring.” Knowledge- (2021): n. pag.
Based Systems, vol. 249, Elsevier BV, Aug. 2022, p. [61] Li, Xiaoli et al. “High fidelity single image blind
109008. Crossref, doi:10.1016/j.knosys.2022.109008. deblur via GAN.” Wireless Networks (2021): 1-10.
[45] Zhang, Jing and Tao Zhang. “Image Deblurring [62] Tomosada, Hiroki et al. “GAN-based image
Algorithm Based on the Gaussian-Scale Mixture deblurring using DCT loss with customized
Expert Field Model.” Mathematical Problems in datasets.” IEEE Access PP (2021): 1-1.
Engineering (2022): n. pag. [63] Zhou, Luoyu and Zhiyang Liu. “Blind Deblurring
[46] Bai, Yang et al. “A Deconvolutional Deblurring Based on a Single Luminance Channel and L1-
Algorithm Based on Dual-Channel Images.” Applied Norm.” IEEE Access 9 (2021): 126717-126727.
Sciences (2022): n. pag. [64] Hsieh, Po-Wen and Pei-Chiang Shao. “Blind image
[47] Parvaz, Reza. “Point spread function estimation for deblurring based on the sparsity of patch minimum
blind image deblurring problems based on framelet information.” Pattern Recognit. 109 (2021): 107597.
transform.” ArXiv abs/2112.11004 (2021): n. pag. [65] Shao, Wenze et al. “Gradient-based discriminative
[48] Zhang, Ziyu et al. “Blind Remote Sensing Image modeling for blind image
Deblurring Using Local Binary Pattern deblurring.” Neurocomputing 413 (2020): 305-327.
Prior.” Remote. Sens. 14 (2022): 1276. [66] Bao, Jiaqi et al. “Half quadratic splitting method
[49] Xia, Haiying et al. “MFC-Net: Multi-scale fusion combined with convolution neural network for blind
coding network for Image Deblurring.” Applied image deblurring.” Multimedia Tools and
Intelligence 52 (2022): 13232 - 13249. Applications 80 (2020): 3489 - 3504.
[50] Sanghvi, Yash et al. “Photon-Limited Blind [67] Zhu, Li et al. “Blind Image Deblurring Based on
Deconvolution Using Unsupervised Iterative Kernel Local Rank.” Mobile Networks and Applications 25
Estimation.” IEEE Transactions on Computational (2019): 1446 - 1456.
Imaging 8 (2022): 1051-1062. [68] Khan, Aftab and Hujun Yin. “Arbitrarily shaped Point
[51] Cao, Shuning et al. “Robust Blind Deblurring Under Spread Function (PSF) estimation for single image
Stripe Noise for Remote Sensing Images.” IEEE blind deblurring.” The Visual Computer 37 (2020):
Transactions on Geoscience and Remote Sensing 60 1661 - 1671.
(2022): 1-17. [69] Bai, Yuanchao et al. “Single-Image Blind Deblurring
[52] Huang, Jun-Jie and Pier Luigi Dragotti. “WINNet: Using Multi-Scale Latent Structure Prior.” IEEE
Wavelet-Inspired Invertible Network for Image Transactions on Circuits and Systems for Video
Denoising.” IEEE Transactions on Image Technology 30 (2019): 2033-2045.
Processing 31 (2021): 4377-4392. [70] Cheng, Shichao et al. “Blind image deblurring via
[53] Liu, Yang et al. “Multi-Scale Grid Network for Image hybrid deep priors modeling.” Neurocomputing 387
Deblurring With High-Frequency Guidance.” IEEE (2020): 334-345.
Transactions on Multimedia 24 (2021): 2890-2901. [71] Fan, Chi-Mao et al. “Selective Residual M-Net for
[54] Askari Javaran, Taiebeh and Hamid Hassanpour. Real Image Denoising.” 2022 30th European Signal

ISSN: 2347-8578 www.ijcstjournal.org Page 141

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
Processing Conference (EUSIPCO) (2022): 469-473. on Dual Attention Network and 2D Blur Kernel
[72] Yang, Dayi et al. “Blind Image Deblurring via a Estimation.” 2021 IEEE International Conference on
Novel Sparse Channel Prior.” Mathematics (2022): n. Image Processing (ICIP) (2021): 1729-1733.
pag. [88] Li, Chen et al. “Multiple Auxiliary Networks for
[73] Biyouki, S. Amrollahi and Hoon Hwangbo. “Blind Single Blind Image Deblurring.” ICASSP 2021 - 2021
Image Deblurring based on Kernel IEEE International Conference on Acoustics, Speech
Mixture.” ArXiv abs/2101.06241 (2021): n. pag. and Signal Processing (ICASSP) (2021): 2070-2074.
[74] Li, Yuelong et al. “Efficient and Interpretable Deep [89] Zhao, Suiyi et al. “FCL-GAN: A Lightweight and
Blind Image Deblurring Via Algorithm Real-Time Baseline for Unsupervised Blind Image
Unrolling.” IEEE Transactions on Computational Deblurring.” Proceedings of the 30th ACM
Imaging 6 (2020): 666-681. International Conference on Multimedia (2022): n.
[75] Ljubenović, Marina and Mário A. T. Figueiredo. pag.
“Blind image deblurring using class-adapted image [90] Zhuang, Zhong et al. “Blind Image Deblurring with
priors.” 2017 IEEE International Conference on Unknown Kernel Size and Substantial Noise.” ArXiv
Image Processing (ICIP) (2017): 490-494. abs/2208.09483 (2022): n. pag.
[76] Tofighi, Mohammad-Reza et al. “Blind Image [91] An, Sungkwon et al. “Blur Invariant Kernel-Adaptive
Deblurring Using Row–Column Sparse Network for Single Image Blind Deblurring.” 2021
Representations.” IEEE Signal Processing Letters 25 IEEE International Conference on Multimedia and
(2017): 273-277. Expo (ICME) (2020): 1-6.
[77] Bai, Yuanchao et al. “Graph-Based Blind Image [92] Fergus, Rob et al. “Removing camera shake from a
Deblurring From a Single Photograph.” IEEE single photograph.” ACM Trans. Graph. 25 (2006):
Transactions on Image Processing 28 (2018): 1404- 787-794.
1418. [93] Cho, Sunghyun and Seungyong Lee. “Fast motion
[78] Zhang, Haichao et al. “Sparse representation based deblurring.” ACM SIGGRAPH Asia 2009 papers
blind image deblurring.” 2011 IEEE International (2009): n. pag.
Conference on Multimedia and Expo (2011): 1-6. [94] Xu, Li et al. “Unnatural L0 Sparse Representation for
[79] Zhang, Hong et al. “Image deblurring using tri- Natural Image Deblurring.” 2013 IEEE Conference on
segment intensity prior.” Neurocomputing 398 Computer Vision and Pattern Recognition (2013):
(2020): 265-279. 1107-1114.
[80] Almeida, Mariana S. C. and Mário A. T. Figueiredo. [95] Zhang, Haichao et al. “Multi-image Blind Deblurring
“New stopping criteria for iterative blind image Using a Coupled Adaptive Sparse Prior.” 2013 IEEE
deblurring based on residual whiteness Conference on Computer Vision and Pattern
measures.” 2011 IEEE Statistical Signal Processing Recognition (2013): 1051-1058.
Workshop (SSP) (2011): 337-340. [96] Xu, Li and Jiaya Jia. “Two-Phase Kernel Estimation
[81] Li, Lerenhan et al. “Learning a Discriminative Prior for Robust Motion Deblurring.” European Conference
for Blind Image Deblurring.” 2018 IEEE/CVF on Computer Vision (2010).
Conference on Computer Vision and Pattern [97] Zhong, Lin et al. “Handling Noise in Single Image
Recognition (2018): 6616-6625. Deblurring Using Directional Filters.” 2013 IEEE
[82] Liu, Jing et al. “Blind Image Deblurring via Local Conference on Computer Vision and Pattern
Maximum Difference Prior.” IEEE Access 8 (2020): Recognition (2013): 612-619.
219295-219307. [98] Krishnan, Dilip et al. “Blind deconvolution using a
[83] Chen, Xiaole et al. “Hyper-Laplacian Regularized normalized sparsity measure.” CVPR 2011 (2011):
Non-Local Low-Rank Prior for Blind Image 233-240.
Deblurring.” IEEE Access 8 (2020): 136917-136929. [99] Michaeli, Tomer and Michal Irani. “Blind Deblurring
[84] Mao, Caisheng et al. “Adaptive Optimization with Using Internal Patch Recurrence.” European
Nested Prior Navigation for Blind Image Deblurring.” Conference on Computer Vision (2014).
2018 IEEE Fourth International Conference on [100] Levin, Anat et al. “Efficient marginal likelihood
Multimedia Big Data (BigMM) (2018): 1-7. optimization in blind deconvolution.” CVPR 2011
[85] Chen, Liang et al. “Blind Image Deblurring With (2011): 2657-2664.
Local Maximum Gradient Prior.” 2019 IEEE/CVF [101] Pan, Jin-shan et al. “Deblurring Face Images with
Conference on Computer Vision and Pattern Exemplars.” European Conference on Computer
Recognition (CVPR) (2019): 1742-1750. Vision (2014).
[86] Nan, Yuesong et al. “Variational-EM-Based Deep [102] Whyte, Oliver et al. “Non-uniform Deblurring for
Learning for Noise-Blind Image Deblurring.” 2020 Shaken Images.” International Journal of Computer
IEEE/CVF Conference on Computer Vision and Vision 98 (2010): 168-186.
Pattern Recognition (CVPR) (2020): 3623-3632. [103] Perrone, Daniel J. and Paolo Favaro. “Total Variation
[87] Tian, Senmao et al. “Blind Image Deblurring Based Blind Deconvolution: The Devil Is in the Details.”

ISSN: 2347-8578 www.ijcstjournal.org Page 142

International Journal of Computer Science Trends and Technology (IJCST) – Volume 11 Issue 2, Mar-Apr 2023
2014 IEEE Conference on Computer Vision and [119] Levin, Anat et al. “Understanding and evaluating
Pattern Recognition (2014): 2909-2916. blind deconvolution algorithms.” 2009 IEEE
[104] Sun, Libin et al. “Edge-based blur kernel estimation Conference on Computer Vision and Pattern
using patch priors.” IEEE International Conference on Recognition (2009): 1964-1971.
Computational Photography (ICCP) (2013): 1-8. [120] Köhler, Rolf et al. “Recording and Playback of
[105] Pan, Jin-shan et al. “Blind Image Deblurring Using Camera Shake: Benchmarking Blind Deconvolution
Dark Channel Prior.” 2016 IEEE Conference on with a Real-World Database.” European Conference
Computer Vision and Pattern Recognition (CVPR) on Computer Vision (2012).
(2016): 1628-1636. [121] Lai, Wei-Sheng et al. “A Comparative Study for
[106] Pan, Jin-shan et al. “L_0 -Regularized Intensity and Single Image Blind Deblurring.” 2016 IEEE
Gradient Prior for Deblurring Text Images and Conference on Computer Vision and Pattern
Beyond.” IEEE Transactions on Pattern Analysis and Recognition (CVPR) (2016): 1701-1709.
Machine Intelligence 39 (2017): 342-355. [122] Pan, Jin-shan et al. “Deblurring Face Images with
[107] Yan, Yanyang et al. “Image Deblurring via Extreme Exemplars.” European Conference on Computer
Channels Prior.” 2017 IEEE Conference on Computer Vision (2014).
Vision and Pattern Recognition (CVPR) (2017): [123] Rim, Jaesung et al. “Real-World Blur Dataset for
6978-6986. Learning and Benchmarking Deblurring
[108] Wen, Fei et al. “A Simple Local Minimal Intensity Algorithms.” European Conference on Computer
Prior and an Improved Algorithm for Blind Image Vision (2020).
Deblurring.” IEEE Transactions on Circuits and
Systems for Video Technology 31 (2020): 2923-2937.
[109] Hirsch, Michael et al. “Fast removal of non-uniform
camera shake.” 2011 International Conference on
Computer Vision (2011): 463-470.
[110] Shan, Qi et al. “High-quality motion deblurring from
a single image.” ACM SIGGRAPH 2008 papers
(2008): n. pag.
[111] Nah, Seungjun et al. “Deep Multi-scale Convolutional
Neural Network for Dynamic Scene Deblurring.”
2017 IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) (2016): 257-265.
[112] Kupyn, Orest et al. “DeblurGAN: Blind Motion
Deblurring Using Conditional Adversarial Networks.”
2018 IEEE/CVF Conference on Computer Vision and
Pattern Recognition (2017): 8183-8192.
[113] Tao, Xin et al. “Scale-Recurrent Network for Deep
Image Deblurring.” 2018 IEEE/CVF Conference on
Computer Vision and Pattern Recognition (2018):
8174-8182.
[114] Zhou, Shangchen et al. “DAVANet: Stereo
Deblurring With View Aggregation.” 2019
IEEE/CVF Conference on Computer Vision and
Pattern Recognition (CVPR) (2019): 10988-10997.
[115] Kupyn, Orest et al. “DeblurGAN-v2: Deblurring
(Orders-of-Magnitude) Faster and Better.” 2019
IEEE/CVF International Conference on Computer
Vision (ICCV) (2019): 8877-8886.
[116] Gao, Hongyun et al. “Dynamic Scene Deblurring
With Parameter Selective Sharing and Nested Skip
Connections.” 2019 IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR)
(2019): 3843-3851.
[117] Qi, Qing et al. “Attention Network for Non-Uniform
Deblurring.” IEEE Access 8 (2020): 100044-100057.
[118] Ye, Minyuan et al. “Scale-Iterative Upscaling
Network for Image Deblurring.” IEEE Access 8
(2020): 18316-18325.

ISSN: 2347-8578 www.ijcstjournal.org Page 143