Volume 6, Issue 2, February – 2021 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Motion Blur Detection and Removal in Images

Alrik Fernandes1, Prathamesh Joshi2, Gaurav Kshirsagar 3, Manav Chordia4
SCTR’s Pune Institute of Computer Technology, Pune, Maharashtra, India

Abstract:- Image blur detection and removal have been Beomseok Kim, Hyeongseok Son, Seong-Jin Park,
one of the major topics of research in image processing Sunghyun Cho, and Seungyong Lee [2] put forward a novel
in the recent years. These blur detection and removal approach for detecting two types of blurs - motion blur and
algorithms have many real world applications including out-of-focus blur. They proposed a method that used a deep
image restoration and image enhancement. Image blur encoder-decoder network with long residual skip-
can include motion blur and out-of-focus blur or connections along with multi-scale reconstruction loss
blurring due to lens imperfections. This paper covers an functions to make use of low-level structural features as well
overview of recent methods and advancements made in as high-level contextual features. This research outperforms
the fields of motion blur detection as well as motion blur other state-of-the-art methods. However this research was
removal. This paper also proposes an approach for performed on a limited size of dataset without considering
motion blur detection and removal involving complex cases of blurs in an image.
Convolutional Neural Network(CNN) and Generative
Adversarial Network(GAN). Karl S. Ni, Zachary Z. Sun & Nadya T. Bliss [3]
proposed an algorithm to detect global motion blur from a
Keywords:- Motion Blur, Out-of-Focus Blur Detection, Blur video source. The working of the algorithm is two fold,
Removal, Image Restoration. firstly the algorithm creates a blur metric from any single
frame or image and secondly, it adds temporal information
I. INTRODUCTION by making use of correlated information from adjacent
frames from a video feed. The advantage of this algorithm is
One of the frequently encountered problems in that it makes use of adjacent reference frames from videos
photography as well as capturing a video is the introduction compared to just blur detection from still images.
of blur either due to object movement or camera motion Additionally it is easy to understand and outputs high
associated with the speed of the camera (shutter speed) performance for blur detection.
when pictures are taken. Blur is the smoothing of the image
pixels essentially resulting in a relatively obscure image. To Bing Li, Zhen Huan Zhan [4] discussed recovering a
counter this problem the initial step is blur detection where degraded image using linear and inverse filter methods.
the part of the image where the blur has occurred is Motion blurred distance and motion blurred direction are the
identified. After this the blur is classified into general blur parameters used in the recovery process of blurred images.
and motion blur. The last step is the image restoration step The authors used mean square error for comparing the
where blur is removed. This paper discusses some quality of image between restored image and blurred image.
approaches used to detect and remove motion blur. This
paper also proposes a deep learning based approach on Taeg Sang Cho [5] discussed a kernel based spatially
detection of blur, targeting mainly motion blur and not a invariant blur detection method that uses blurred-edge
plain defocused image by using a culmination of different profiles. The research proposed both a hardware and a
methodologies and once detected, work on remediating it software solution for blur detection and it’s removal. The
using Generative Adversarial Network in order to restore it software based solutions involve blur kernel estimation
or rather regenerate the original image. through blurred line profiles and using phase information
while the hardware solution involves using a camera that
II. LITERATURE SURVEY improves local motion estimation using computations.

Renting Liu, Zhaorong Li & Jiaya Jia [1] proposed a Dong Gong, Jie Yang [6] proposed a pixel wise linear
partial-blur image detection along with a framework that motion blur representation for heterogeneous motion blur.
analyses and classifies types of blurs without the need of The method the authors proposed estimates a dense motion
deblurring. They used several blur features derived by colors flow map using a fully-convolutional deep neural network.
from an image along with it’s spectrum & gradient The dataset they used consists of both real world and
information & used these feature parameters to robustly synthetic images. The research shows promising results for
train & classify blurred images. The blur is classified as real world images containing heterogeneous motion blur.
either motion blurred regions caused due to motion of
subject or out-of-focus blur which is generally caused by
lenses being out-of-focus from subject in an image.

IJISRT21FEB187 www.ijisrt.com 450

Volume 6, Issue 2, February – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Shuang Zhang, Ada Zhen [7] put forward an approach III. PROPOSED METHODOLOGY
wherein a pair of images as input are subject to denoising
and deblurring encoders after which they are merged and A. Detection
passed through a deblurring decoder. The two encoders are In this section, we propose an end-to-end deep neural
used to extract the complimentary information from the pair network for detecting motion blur and localizing it. The
of images. The merger is done to combine the information in problem of motion blur can be assumed as a problem of
parallel. However a prerequisite for their model are pairs of semantic segmentation with the regions containing motion
blurry images which is a drawback. blur to be segmented. We propose the use of a Mask region-
based convolutional neural network for this task. A Mask R-
Jian Sun, Wenfei Cao, Zongben Xu & Jean Ponce [8] CNN is a convolutional neural network aimed to solve the
put forward a novel deblurring approach based on instance segmentation problem. There are two stages in
convolutional neural networks for images containing non- Mask R-CNN. First, the region proposal network generates
uniform motion blur. This deep learning approach is used proposals for the regions where an object might be present
for predicting probabilistic distribution of the motion blur at based on the input image. Second, it classifies the object,
the patch level. A Markov random field is used to enforce refines the bounding box, and generates a pixel level mask
motion smoothness after which the motion blur is removed of the object based on the first stage proposal. [See fig. 1]
using a non-uniform deblurring model.
 Homogenous regions:
Jian-Feng Cai, Hui Ji, Chaoqiang Liu and Zuowei Some flat regions having the same colour are
Shen [9] discussed a new approach to remove motion blur sometimes misclassified as blur in order to remediate this,
that is based on high sparsity of the motion blur kernel in the we separately process the image as separate blocks of sizes
curvlet system and that of the image in the framelet system. 8x8 or 16x16 and then plot the standard deviation of the
The method requires no prior information on the kernel spectral intensities of the pixels. If the standard deviation is
which makes it different from existing approaches. The very low then we can understand that the region is
algorithm was tested extensively on synthesized as well as essentially black and that there is no motion blur present the
real world images and shows promising results compared to Intersection over Union is then taken for the combining the
existing approaches. outputs of the standard deviation mask and the mask rcnn
output.Essentially this makes our model more robust to any
kind of image and improves the performance of the motion
blur detection .

Fig 1:- Architecture diagram of proposed methodology for blur detection [10], [11]

B. Removal 4. The image generated by the GAN generator is sent to the

In this section, we propose an end-to-end deep neural discriminator to calculate loss.
network for removing motion blur. This problem of 5. Based on the loss function generator and discriminator
removing motion blur can be assumed as a problem of loss is calculated and back propagation is performed.
deconvolving the PSF that leads to the blur. We propose the
use of a Generative Adversarial Network for this task. IV. CONCLUSION
GANs are an approach to generative modeling using deep
neural networks. GANs will be used for training a This paper starts with the introduction of motion blur
generative model that will generate deblurred images. and its applications. Several motion blur detection and
removal techniques were introduced. We also discussed
 Algorithm approaches involving a single image as an input as well as a
1. Train a custom Mask R-CNN[12] for semantic pair of images. We also discussed how encoders and
segmentation between non-blurry and blurry portions of decoders are effectively used to detect and remove motion
the image. blur. Lastly we proposed an approach involving a mask
2. Based on the pre-decided threshold, decide if the image region-based convolutional neural network for detection and
is blurry or not. If blurry send to blur removal. a Generative Adversarial Network based method for
3. For blur removal train a GAN[13] with sharp and removal.
respective motion blurred images.

IJISRT21FEB187 www.ijisrt.com 451

Volume 6, Issue 2, February – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
ACKNOWLEDGEMENTS [13] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi
Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
It gives us great pleasure and satisfaction in presenting Aaron Courville, Yoshua Bengio, “Generative
this paper titled “Motion blur detection and removal in Adversarial Networks” in “Proceedings of the 27th
images”. International Conference on Neural Information
Processing Systems - Volume 2”, 2014
Our special thanks to Dr. A.M. Bagade for their [14] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi
support and guidance. Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, Yoshua Bengio, “Generative
We would also like to thank external guide Mr. Adversarial Networks” in “Proceedings of the 27th
Pushkar Nimkar and reviewers Mr. R.B. Murumkar and International Conference on Neural Information
Sarika U. Patil for their time to time feedback and guidance. Processing Systems - Volume 2”, 2014
[15] Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He,
We would like to thank the Department of Information Jianfeng Gao, “MS-Celeb-1M: A Dataset and
Technology, PICT and all those, who have directly or Benchmark for Large-Scale Face Recognition”, 2016
indirectly helped us for the completion of the work during [16] Gary B. Huang, Manu Ramesh, Tamara Berg, and
this paper. Erik Learned-Miller. Labeled Faces in the Wild: A
Database for Studying Face Recognition in
REFERENCES Unconstrained Environments.

[1] Renting Liu, Zhaorong Li & Jiaya Jia, “Image partial

blur detection and classification” in CVPR 2008.
[2] Beomseok Kim, Hyeongseok Son, Seong-Jin Park,
Sunghyun Cho, and Seungyong Lee, “Defocus and
Motion Blur Detection with Deep Contextual
Features” in Computer Graphics Forum 37 (7) (2018)
[3] Karl S. Ni, Zachary Z. Sun & Nadya T. Bliss, “Real-
time global motion blur detection” in 9th IEEE
International Conference on Image Processing, 2012
[4] Bing Li, ZhenHuan Zhan, “Research on Motion
Blurred Image Restoration,” in 5th International
Congress on Image and Signal Processing, 2012.
[5] Taeg Sang Cho, “Motion Blur Removal from
Photographs ”, in . M.I.T Ph.D dissertation (2010)
[6] Dong Gong, Jie Yang, “From Motion Blur to Motion
Flow:a Deep Learning Solution for Removing
Heterogeneous Motion Blur”, CVPR, 2017
[7] Shuang Zhang, Ada Zhen, “Deep Motion Blur
Removal Using Noisy/Blurry Image Pairs”, Cornell
University, 2019
[8] Jian Sun, Wenfei Cao, Zongben Xu, Jean Ponce,
“Learning a Convolutional Neural Network for Non-
uniform Motion Blur Removal” , CVPR, 2015
[9] Jian-Feng Cai, Hui Ji, Chaoqiang Liu and Zuowei
Shen, “Blind motion deblurring from a single image
using sparse approximation”, CVPR, 2009
[10] Xuewei Wang, Shulin Zhang, Xiao Liang, Hongjun
Zhou, Jinjin Zheng, Mingzhai Sun, “Accurate and
Fast Blur Detection Using Pyramid M-Shaped Deep
Neural Network”, IEEE, 2019
[11] Wenjun Yu, Sumi, Fei Chen, Jaeho Choi, “Pedestrian
Detection Based on Improved Mask R-CNN
Algorithm” , Springer, 2020
[12] Ross Girshick, Jeff Donahue, Trevor Darrell, Jitendra
Malik, “Rich feature hierarchies for accurate object
detection and semantic segmentation” in CVPR 2013

IJISRT21FEB187 www.ijisrt.com 452