ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 3, September 2012
Image Deblurring Using a Neural Network
Approach
Neetin Kumar, Dr. Manish Shrivastava, Naazish Rahim
Abstract— This paper presents the restoration of blurred II. RESTORATION USING A NEURAL NETWORK
image which is degraded using blind disconsolation algorithm.
Consider the single layer perceptron in figure 1
There are many researches done using this algorithm [6], but no
Single Layer Perceptron.
one conceived it from parallel computing procedure. In this
work, we consider the problem of restoring a blurred image
INPUT IMAGE MOCK DEGRADED IMAGE
degraded by blur and high impulsive noise levels. Deblurring
and denoising are probably two of the most studied problems in
image processing. We present a novel algorithm for blind image
deblurring from a single image using parallel computation. The
blur point spread function (PSF) is assumed uniform. We divide
the image and exert the algorithm on each part parallelly. Before
applying the blind image deconvolution algorithm in blurred
image we train the network using back propagation algorithm,
for number of iterations used in deblurring process to finding
the true PS.
Index Terms— Blind Image Deconvolution, Degradation,
Point Spread Function (PSF). BLUR FUNCTION
I. INTRODUCTION Fig 1. Single Layer Perceptron
If the input layer represents an input image, then by
Image restoration is the problem of finding an estimate of
appropriately assigning weights to mimic the local
an ideal image from its blurred and noisy rendition. In many
interactions between neighboring pixels, the output would be
cases these degradations cannot be avoided, and the data
a blurred version of the input image. We shall refer to this as
must be recovered from the post processing. However, most
the mock degraded image. Thus the mock image should
of current digital image processing techniques are always
closely resemble the given degraded image if the input were
assumed that all the information which is required to restore
the true original image (just how close is subject to noise),
image is known a priori [7][8]. Image restoration is to
and any deviation of the input trial image from the original
improve the quality of the degraded image. It is used to
will manifest itself in the error between the mock and the
recover an image from distortions to its original image. It is
actual, observed degraded image. The current trial input can
an objective process which removes the effects of sensing
therefore be improved by back propagating the error signal
environment. It is the process of recovering the original
through the network using the following modified delta rule
scene image from a degraded or observed image using
Δfin = -η (dE/dfin-1)+αΔ fin-1 (1)
knowledge about its nature. Blind Image Deconvolution is a
Where E is the error signal (or more conventionally referred
more difficult image restoration where image recovery is
to as the energy function), and the upper index represents the
performed with little or no prior knowledge of the degrading
number of iteration [9]
PSF. The advantages of Deconvolution are higher resolution
A. Quantization
and better quality [4]. This paper is structured as follows:
The image is partitioned into smaller frames. The whole
Section 2 represents restoration using a neural network.
image will contain same information so this part will make
Section3 Quantization of an image. Section 4 describes the
algorithm computationally efficient and suitable for parallel
degradation model for blurring an image. Section 5 describes
computation [6]. Image Model for kernel assessment
the blurring concept algorithm and. Section 6 overall
architecture of this paper. Section 7 describes the sample assuming uniform blur distribution avoiding non linearity.
results for deblurred images using our proposed algorithm. a.G(x, y) =k(x, y)*f(x, y) + n(x, y) (2)
Section 8 describes the conclusion, comparison and future B.Discrete Fourier transform can be used to yield
work. frequency domain model
G(m,n) = K(m,n)*F(m,n)+N(m,n) (3)
Where m and n discrete horizontal and vertical frequency
variable.
155
ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 3, September 2012
c. Inverse procedure will compute A. Blurring Types
G(m,n)K-1(m,n)=F(m,n)+N(m,n)k-1(m,n) (4) In digital image there are 3 common types of Blur effects:
The following model, after partition holds approximately: 1) Average Blur.
Gi (x,y)=K(x,y)Fi (x,y)+N i (x,y);i = 1,2,3…n p (5) 2) Gaussian Blur.
Where np is the total number of partitioned subsections. 3) Motion Blur.
A) Algorithm for Image Quantization[6]
III. OVERALL ARCHITECTURE
Input: The following Fig. 2 represents the overall architecture of
Load an input image „I‟. this paper .The original image is degraded or blurred using
Initialize row value. degradation model to produce the blurred image. The blurred
Initialize column value. image should be an input to the deblurring algorithm.
Procedure – I Various algorithms are available for deblurring. In this
Load an input image I paper, we are going to use blind deconvolution algorithm.
Partition the input image I The result of this algorithm produces the deblurring image
G i (x,y) = K (x,y) F i (x,y) + N i (x,y); i = 1,2,3……n p which can be compared with our original image.
Go to Procedure – II
End Procedure – I
B.Degradation Model
In the model of image degradation [5] fig. 1, the blurred
image can be characterized by its degradation function, the
noise function and by original image.
1) Blind Deconvolution Algorithm:
Blind Deconvolution Algorithm can be used effectively
when no information of distortion is known. It restores image
and PSF simultaneously. Definition of the blind deblurring
B) Algorithm for Degradation Model method can be expressed by:
Input: G(x, y) =h(x, y)*f(x, y) + n(x, y) (6)
Load an input image „I‟ Where: g (x, y) is the observed image, f(x,y) is Point
Initialize blur length „l‟
Spread Function, h(x,y) is the constructed image and η (x,y)
Initialize blur angle „theta‟ is the additive noise term .
Assign the type of noise „n‟ C) Algorithm for Deblurring each frame:
PSF (Point Spread Function), „F‟ Input:
Procedure – II Back propagation Algorithm‟BA‟
h=create (I, l, theta) %Creation of PSF Blurred image „B‟
Blurred image (B) = I*F + n Initialize number of iterations „j‟
B= filter (I, f, n,”convolution”) Initial PSF „F‟
Go to Procedure – III Procedure – III
End Procedure – II If PSF is not known then
C. Blurring Guess initial value of PSF
Blur is unsharp image area caused by camera or subject Else
movement, inaccurate focusing, or the use of an aperture that Specify the PSF of degraded image
gives shallow depth of field. The Blur effects are filters that Apply back propagation algorithm BA,
smooth transitions and decrease contrast by averaging the
Restored Image I‟= Deconvolution (B, F, j,Gi(x,y))
pixels next to hard edges of defined lines and areas where Combine all the restored images
there are significant color transition. End Procedure – III
156
ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 3, September 2012
Blurred Image
IV. SAMPLE RESULTS Blurred Image
The, Fig 4(a) the original image.We break the original
Image and apply different PSF on every part named A,B,C,D
as shown in Fig 5 (a)(b)(c)(d). Fig 6 (a)(b)(c)(d) is depicted as
blurred images for all parts A,B,C,D. We apply the neural
network approach , back propagation algorithm for each
iteration of PSF. Fig 7 shows the trained output of first Fig 6(A) (B) (C) (D) Four Divided Blurred Image
iteration using back propagation, and Fig 8 shows Trained 1
10
Performance is 0.513057, Goal is 0
output after final iteration using back propagation algorithm
The sample image after applying the proposed algorithm will
be shown in Fig. 9. The result of our algorithm is quite
Training-Blue
0
10
promising from parallel computation view. We can compute
the each part independent from each other and final result
will be combination of all divided images.
-1
original image 10
0 10 20 30 40 50 60 70 80 90 100
100 Epochs
Fig 7 Trained Output First Iteration Using Back
Propagation Algorithm
-1
Performance is 0.00442531, Goal is 0
10
Training-Blue
-2
10
-3
10
0 10 20 30 40 50 60 70 80 90 100
100 Epochs
Fig 8 Trained Output after Final Iteration Using
Back Propagation Algorithm
deblur image
Fig 4 (a) Original Image
Fig 5. (a)(b)(c)(d) Four divided Original Image Fig 9 Final Image Computed From Our Algorithm
Blurred Image
Blurred Image
VI. CONCLUSION AND FUTURE WORK
This paper addresses the blind image restoration problem,
on single image using unknown PSF, one wants to restore the
original image. A new model is proposed for this restoration,
the image is divided into different frames and the algorithm
is applied to each frame parallelly also neural network
157
ISSN: 2277-3754
ISO 9001:2008 Certified
International Journal of Engineering and Innovative Technology (IJEIT)
Volume 2, Issue 3, September 2012
approach i.e back propagation algorithm is used for
iterations of different PSF. The restoration problem at hand
in each of these approaches reduces to the problem of solving
a multiple independent set of images on which algorithm is
applied parallel. The future work of this paper is to increase
the speed of the deblurring process that is reducing the
number of iteration using for deblurring the image for
achieving better quality image and by training the network it
will take an advantage of to detect an appropriate image.
.
REFERENCES
[1] Michael Elad and Arie Feuer, Senior Member, IEEE
Restoration of a Single Super resolution Image from Several
Blurred, Noisy, and Under sampled Measured Images.
[2] WANG Shoujue, CAO Yu, HUANG Yi A Novel Image
Restoration Approach Based on Point Location in
High-dimensional Space Geometry.
[3] Rajeev Srivastava*, Harish Parthasarthyt, JRP Guptat and D.
Roy Choudharyl Image Restoration from Motion Blurred
Image using PDEs formalism.
[4] Ms.S.Ramya and Ms.T.Mercy Christial”Restoration of
blurred images using Blind Deconvolution Algorithm”
[5] D. Kundur and D. Hatziankos, “Blind image
deconvolution, “IEEE sig. process. Mag., pp. 43-64, May
1996.
[6] H.C. Andrews, B.R. Hunt, Digital Image Restoration,
Prentice-Hall, Inc., 1977.
[7] M.I. Sezan, A.M. Tekalp. 'Tutorial review of recent
developments in digital image restoration," Proc. SPIE
Visual Commu. Image Processing, vol. 1360, pp.
1361359, 1990.
[8] Chong Sze Tong” Double Deconvolution using a Neural
Network”, proceeding of 1994 International Symposium on
Speech, Image Processing and Neural Networks, 13-16
April 1994, Hong Kong.
158
