multiple inpainting of low resolution images using...
TRANSCRIPT
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3641
Multiple Inpainting of low resolution images using
examplar and super resolution algorithm
A.Breezy Joy, C.Rekha
Abstract—Inpainting is the process of filling the missing regions
in an image. The main aim of this paper is to fill the missing areas
using examplar based inpainting and to recover the missing areas
and improve the quality of the image using super resolution
algorithm. The performance of the algorithm is evaluated using
PSNR, mean square error and Histogram error. The damaged
image is first downsampled. Then the image is inpainted a
number of times using examplar based approach. The inpainted
images are combined using loopy belief propagation. The image is
finally upsampled and the image quality is improved using a
super resolution algorithm.
Index Terms—examplar based inpainting, mean square error,
loopy belief propagation, running time.
1. INTRODUCTION
Inpainting is used for reconstructing lost or
deteriorated parts of images. Image inpainting techniques are
in use over a long time for various applications like scratch
removal, restoring the damaged, missing portions or removal
of objects from the images, etc. In photography and cinema,
inpainting is used for film restoration. To reverse the
deterioration, dust spots in film and see infrared cleaning. It is
also used for removing red-eye, eliminating the stamped date
from photographs and removing objects for creative effect.
This technique is be used to replace the lost blocks in the
transmission and coding of images such as in a streaming
video. It is also used to remove the logos in videos. If any
unwanted portion of the image is to be removed, then the
missing region can be filled using inpainting.
In a paper fragment based image completion,
adaptive neighborhood algorithm is used. Here the most
similar fragments from the neighborhood are selected and
filled in the missing regions. The visible part of the image
serve as a training set for the unknown parts. But in this
method, more time is spent for searching the best matching
fragments. Sometimes two fragments of same part may differ
under slight illumination changes. The next paper is
inpainting by patch propagation using patch sparsity. In this
method the missing patches are represented by sparse linear
combination of candidate patches. The patch with large
structure sparsity is used for inpainting. Here only the
structure of the patch is considered. Texture is not included.
Examplar based inpainting based on local geometry
uses K-nearest neighbor algorithm to find the structure tensor.
To find the best matching patch, template matching is
performed. This method fails to replicate texture in a coherent
manner. In texture synthesis method, the textural information
and structural information are propagated and the patch
priority is calculated. This method fails to handle curved
structures. If similar patches do not exist, it does not produce
reasonable output.
In a paper statistics of patch offsets for image
completion, the statistics of offsets which are sparsely
distributed are obtained. Some dominant offsets provide
reliable information for filling the image. A stack of shifted
images are combined by optimization and it is used to fill
missing region. Dominant offsets show how the patterns are
repeated and give details about the missing regions. The main
advantage is this method has better results and has fast image
completion. But this method fails when the desired offset does
not have dominant statistics. This problem can be solved only
partially by introducing manual offsets.
In non parametric sampling, a markov random field
model is assumed and the conditional distribution of each
pixel is calculated. A number of textures are synthesized from
a single texture. But some textures may slip into wrong part
and produce verbatim copies of the original. Only frontal
parallel textures are handled. Local linear embedding is used
to generate a high resolution image from a low resolution
image. Here more number of neighbours should be considered.
All these above methods have low signal to noise ratio and
image quality is low. These drawbacks are overcome by the
proposed method.
The proposed method uses examplar based
inpainting. Inpainting is performed on a coarse version of the
input image. For filling the missing regions a non parametric
patch sampling method is used. In this method the best
matching patches from the neighborhood are chosen and they
are filled in the missing region.To recover fine details on the
missing areas, super resolution algorithm is used which
converts the low resolution image back to high resolution
image.Template matching, non local means and non negative
matrix factorization. These three methods aimed at selecting
the best matching pixel which is to be replaced in the place of
inpainting region. The parameters like PSNR, MSE and
Histogram error are estimated.
2. PROPOSED METHOD
An image to be inpainted is taken. As low resolution
images are easier and efficient for inpainting, the image is first
downsampled. i.e., a high resolution image is converted into
low resolution image. Then the image is inpainted for a
number of times using examplar based inpainting. Then all
the inpainted images are combined using loopy belief
propagation. Then to improve the quality of the image, a super
resolution algorithm is applied on the image to improve image
quality and then convert the low resolution image back to high
resolution images. Thus the final inpainted image with high
resolution is obtained and the performance is evaluated.
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3642
Fig.2.1Block diagram for inpainting
3. EXAMPLAR BASED INPAINTING
The first step in image inpainting is to convert the
high resolution image into low resolution image into low
resolution image. For this converting process a Gaussian
pyramid decomposition is performed. A low pass filtering
operation is performed on the image pixels. Downsampling is
necessary because the low resolution images are easier to
inpaint. Low resolution images are contaminated by noise.
Local orientation singularities which affect the filling process
are strongly reduced if low resolution images are used.
After converting into low resolution images, mark the
target region to be filled using region of interest. The
remaining part of the image is the source region from which
the neighbouring patches are selected.
I
A Boolean matrix is constructed which stores a
binary value 1 for pixel to be inpainted and binary value 0 for
the other pixels. This matrix is called fill region. The next step
is to find the boundary of marked region by convolving
the fill region with laplacian filter. The next step is to find the
priority of the patch which is to be filled first. Texture
synthesis is performed which copies the samples from the
neighbourhood and replaced in the missing region.
3.1 FIND PATCH PRIORITY
The missing patch which has the highest priority is to
be found. Priority is the patch which has less unknown
regions. The patch priority is given by the product of data term
and confidence term.
P(p) = C(p) D(p)
Where P(p) is the patch priority
C(P) is the confidence term
D(p) is the data term
Data term can be either a tensor based or sparsity based data
terms. Tensor based priority term is given by
m
i
TIiIiJ1
Where J is the sum of structure tensors of each scalar image
channels. The scalar structure tensor undergoes smoothing
operation by
GJJ
Where G is the Gaussian distribution with standard deviation
. A similarity weight is measured between the known
patches and the unknown patch. For any selected patch, a
collection of neighbouring patches with highest similarities are
distributed in the same structure or texture. The confidence
values of structure for a patch are measured by the sparseness
of its non-zero similarities to the neighbouring patches. The
patch which is more sparsely distributed with non zero
similarities is placed on the fill front due to more structure
sparseness. For a patch located at the fill front , a
neighborhood window N (p) is set with centre p.
3.2 TEXTURE SYNTHESIS
Texture synthesis is the process of algorithmically
constructing a large digital image by taking advantage of the
structural content. Texture synthesis is used for filling the
missing parts. The patch which has the highest priority is filled
first. To fill a patch, the most similar patch from the remaining
part of the image is taken. A similarity metric is used to find
the best matching patch.
Texture synthesis is of two types. Pixel based texture
synthesis and patch based texture synthesis. In pixel based
method, a texture is synthesized in scan line order by finding
and copying pixels with most similar local neighborhood as
the synthetic texture. This method is very useful for image
completion.
IMAGE
N
Inpainting 1
Inpainting2
Inpainting 3
Belief
propagation
SR
algorithm
Enhanced
image
Inpainting n
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3643
In patch based texture synthesis, a new texture is
created by copying and stitching the textures at various offsets.
This method is more effective and faster than pixel based
texture synthesis. The chosen patch should have maximum
similarity between the known pixel values of the current patch
to be filled and the co-located pixel values.
Coherence measure is given by
)),((min)(jx
Sjx ppSSD
pp dcoh
dSSD- sum of square differences
Ψpx - current patch
Ψpj – neighbouring patch
Coherence measure is the degree of similarity
between the synthesized patch and the original patch. The
chosen neighbor should lie within min)1( d where
mind is the minimum distance between the current patch and
the closest neighbor.
If the selected patch does not satisfy this
condition, then the filling process is stopped and the priority of
the current patch is decreased. Again the patch with highest
priority is chosen. The unknown parts are pasted using the best
matching patch by direct sampling. Alpha blending is used to
combine the known part and the source patch. Alpha blending
is the process which displays a bitmap which has transparent
or semi transparent pixels. The image is made transparent and
the known part and unknown part are combined together.
Poisson fusion is the process used to hide the seams. The
unevenness caused in the image in the inpainting process is
corrected by using Poisson fusion. Patch size and filling order
must be considered. Patch size can be 55, 77, 99, 1111 etc. a parameter M is considered which is the number of
times inpainting is done.
3.3 LOOPY BELIEF PROPAGATION
LBP is a message passing algorithm. A node is used to pass a
message to the adjacent node only when it has received all the
messages, eliminating the message from the destination node
to itself. In the same way, all the inpainted images are
combined together using loopy belief propagation. A label is
assigned to each pixel of the unknown regions T of the image.
A major drawback of belief propagation is that it is slow when
the number of labels is high. So loopy belief propagation is
used to avoid this complexity.
To solve the problems like blur and spatial
consistency, there is a need to minimize the objective function.
The number of labels assigned must be equal to the number of
patches in the source region. In loopy belief propagation, the
number of labels is small. Label is the index of the inpainted
image from which the patches are extracted. A label is
assigned for each pixel so that the total energy E of the markov
random field is minimized.
Energy of MRF is given by
E(l)=∑ vd(lp) + ∑ Vs(ln , lm)
vd (lp) is the label cost
Vs (ln , lm) is the discontinuity cost
The cost increases when the similarity between the current
patch and collocated patches are less. Discontinuity cost is the
difference between labels. λ is the weighing factor and it is set
to 100.Using loopy belief propagation, minimization of energy
E is performed over the target region T and it corresponds to
maximum a posteriori (MAP) estimation problem. When λ is
0, there is no smoothness term. Some artifacts are visible. A
good trade-off is obtained by setting the value of λ to100.
4. SUPER RESOLUTION ALGORITHM
After combining all the inpainted images, a super
resolution algorithm is used to recover back the high
resolution image from the low resolution image. The
downsampled image is upsampled. It is also used to improve
the quality of the image. In low resolution image, the pixel
density within an image is small. Hence it offers fewer details.
In a high resolution image, the pixel density is larger. Hence it
offers more details. Super resolution is the process of
obtaining a high resolution image to low resolution image. To
increase the image resolution, the pixel size is reduced by
increasing the number of pixels per unit area. But the amount
of light available per pixel also decreases. The chip size can be
reduced but the increase of capacitance leads to storage
problem. SR image reconstruction is computationally effective
and cost is less.
5. ALGORITHM FOR INPAINTING
Read an image and find the size of the image. Using
median filter, filter the image which replace the pixel
value with the neighbourhood of its value.
Convert the RGB colour image into gray scale image.
Create a mask which is completely filled with 0 and
the area to be inpainted is marked as 1.
Create a label matrix which indicates the label for the
connected components and used to mark the part of
region to be inpainted.
To set the boundary (Ymax, ymin, Xmax, Xmin) in the
inpaint image, compute the values using the below
mentioned equations
1. Ymin=min(row)-1
2. Ymax=max(row)+1
3. Xmin=min(column)-1
4. Xmax=max(column)+1
Calculate the nearest matching pixel in column and
row,
Matching pixel row, A= (Ymax-Ymin)+1
Matchingpixelcolumn, B= (Xmax-Xmin) +1
Find the bottom, top, left and right area of the image,
1. Bottom Area Range (BAR) & Bottom Mean
If (Ymax+A) <= width of the image, then BAR=
(Ymin+A) to (Ymax+A) in x direction
=Xmin to Xmax in y direction.
Otherwise,
BAR = (Ymin+A) to width of image in x direction
=Xmin to Xmax in y direction.
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3644
∑1, 2….n pixels/ (number of pixels) =mean (BAR)
2. Top area range (TAR) & top mean
If (Ymin-A) >= 1, then
TAR= (Ymin-A) to (Ymax-A) in x direction
TAR=Xmin to Xmax in y direction.
Otherwise,
TAR=1 to (Ymax-A) in x direction
=Xmin to Xmax in y direction.
∑1, 2….n pixels/ (number of pixels)=mean(TAR)
3. Right area range (RAR) & right mean
If (Xmax+B) <= height of the image then,
RAR=Ymin to Ymax in x direction
=(Xmin+B) to (Xmax+B) in y direction.
Otherwise,
RAR=Ymin to Ymax in x direction
=(Xmin-B) to height of image in y direction.
∑1,2….n pixels/(number of pixels)=mean(RAR)
4. Left area range (LAR) & right mean
If (Xmin-A) >= 1, then,
LAR=Ymin to Ymax in x direction
=(Xmin-B) to (Xmax-B) in y direction.
Otherwise,
LAR=Ymin to Ymax in x direction
=1 to (Xmax-B) in y direction.
∑1,2….n pixels/(number of pixels)=mean(LAR)
In template matching method, the maximum of the
mean values of bottom, top, left and right area are
calculated. The area to be inpainted is replaced by the
maximum of the above four mean values.
In NLM method, if (bottom mean-top mean) is
greater than 20, assume the top to bottom value as 0,
otherwise 1.
If (left mean-right mean) is greater than 20, then the
left to right value is 0, otherwise 1.
If both top to bottom and left to right values are 1,
then the inpainting area is replaced by,
pixel=mean(bottom,top,right,left);
If top to bottom value is 1 and left to right value is 0
then the inpainting area is replaced by,
Pixel=mean (bottom_mean, top_mean)
If the top to bottom value is 0 and left to right value is
1, then the inpainting region is replaced by
Pixel=mean (right_mean,left mean)
If both top to bottom value and left to right are 0, and
(left mean-right mean) > (bottom mean-top mean),
then the inpainting region is replaced by,
Pixel=mean (bottom_mean, top_mean)
Otherwise the pixels are replaced by pixel=mean
(right_mean, left_mean)
In NMF method, consider a temporary value. If the
miss image is less than 10, the inpainting region is
replaced by
Missing pixel= max (temporary value)
Finally the image has to be enhanced. Median filter is
used to filter the images obtained from the three
methods and hence the image quality is increased.
The performance of inpainting is evaluated by
calculating PSNR mean square error and histogram
error.
6. RESULTS AND DISCUSSION
The final output consists of the input image, inpainted
image, TM image, NLM image and NMF image. First the
input image is displayed. Next the binary mask is displayed in
which the region to be inpainted is in white colour and the
remaining portion is displayed as black. The region to be
inpainted has the pixel value 1 and the remaining part has the
pixel value 0. Three methods of inpainting are performed and
their outputs are displayed. The image quality is improved and
the downsampled image is converted back to high resolution
image. The three methods include Template matching (TM),
non local means (NLM) method and non negative matrix
factorization (NMF) method. Output of each method is
displayed.
Then signal to noise ratio is measured for the three
images. But it does not have much variation. Next mean
square error is calculated. MSE also does not have
considerable changes among the three images. Finally
histogram error is calculated. Histogram error is less in TM
image than the other two images. Hence experimental results
show that among the three methods, TM image is better. Three
parameters, signal to noise ratio, mean square error and
Histogram error are estimated to evaluate the performance of
the three methods. Template matching is better than the other
two methods.
Four examples are considered. For baby image, leaf
image, nature image and blue image TM, NLM and NMF
methods of inpainting are performed. The table 6.1, shows that
the PSNR and MSE values are nearly equal and does not show
much changes for the three methods. To find which method is
better we have to go for another parameter. Histogram error is
calculated for TM, NLM and NMF methods. Histogram error
is less for TM image as compared to the other two methods.
Template matching is the process of selecting the best
matching pixel by taking the mean of the top, bottom, left and
right value. As the inpainting area is replaced by the mean of
the pixel, the inpainting process will be better. Hence TM
method is suitable for inpainting. The comparison of PSNR,
MSE and Histogram error is shown in the table below. The
output figures of each method with different examples are
shown below.
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3645
6.1 Comparison of PSNR, MSE and Histogram Error
Image name Method PSNR MSE Histogram Error
Baby
TM 49.3500
0.7552
2.7458e-005
NLM 49.2528
0.7723
3.1116e-005
NMF 49.3539
0.7545
2.8679e-005
Leaf TM 30.4311
26.8806
2.3023e-004
NLM 30.4109
26.1553
3.1217e-004
NMF 30.4445
26.6987
3.2135e-004
nature TM 38.0800
10.1177
8.7285e-005
NLM 38.0730
10.1340
8.9937e-005
NMF 38.0841
10.1082
9.2380e-005
blue TM 36.8829
13.3287
2.1635e-004
NLM 36.8840
13.3254
2.9234e-004
NMF 36.8841
13.3252
2.4539e-004
Table 6.1 Comparison of PSNR, MSE and Histogram Error
Original image damaged image inpainting region
TM image NLM image NMF image
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3646
Input image damaged image inpainting region
TM image NLM image NMF image
Input image Damaged image Inpainting area
TM image NLM image NMF image
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3647
Input image Damaged image Inpainting area
TM image NLM image NMFimage
7. CONCLUSION
Thus the inpainting was done for a series of images
using examplar based inpainting First the input image is
downsampled to make the inpainting easier. Then inpainting is
done for a number of times using three methods of inpainting.
They are template matching (TM), non local means (NLM)
and non negative matrix factorization (NMF). Finally a super
resolution algorithm is used to improve the quality of the
image. It undergoes a filtering operation to reduce the noise in
the image. Among the three methods, the PSNR and mean
square error does not show much variation. But histogram
error is low for TM image as compared to the other two
methods. Hence template matching is the best method suitable
for inpainting. Future work would be to extend this to other
super resolution methods and also to improve the image
quality after inpainting.
REFERENCES
[1] M. Bertalmio, G. Sapiro, V. Caselles, and C. Ballester,
―Image inpainting,‖ in Proc. 27th Annu. Conf. Comput.
Graph. Interact. Tech., Jul. 2000, pp. 417–424.
[2] D. Tschumperlé and R. Deriche, ―Vector-valued image
regularization with PDEs: A common framework for
different applications,‖ IEEE Trans. Pattern Anal. Mach.
Intell., vol. 27, no. 4, pp. 506–517, Apr. 2005.
[3] T. Chan and J. Shen, ―Variational restoration of non-flat
image features: Models and algorithms,‖ SIAM J. Appl.
Math., vol. 61, no. 4, pp. 1338–1361, 2001.
[4] A. Criminisi, P. Pérez, and K. Toyama, ―Region filling
and object removal by examplar-based image inpainting,‖
IEEE Trans. Image Process., vol. 13, no. 9, pp. 1200–1212,
Sep. 2004.
[5] I. Drori, D. Cohen-Or, and H. Yeshurun, ―Fragment-
based image completion,‖ ACM Trans. Graph., vol. 22, no.
2003, pp. 303–312, 2003.
[6] P. Harrison, ―A non-hierarchical procedure for re-
synthesis of complex texture,‖ in Proc. Int. Conf. Central
Eur. Comput. Graph., Vis. Comput. Vis., 2001, pp. 1–8.
[7] C. Barnes, E. Shechtman, A. Finkelstein, and D. B.
Goldman, ―Patch- Match: A randomized correspondence
algorithm for structural image editing,‖ ACM Trans. Graph.,
vol. 28, no. 3, p. 24, Aug. 2009.
[8] A. A. Efros and T. K. Leung, ―Texture synthesis by non-
parametric sampling,‖ in Proc. 7th IEEE Comput. Vis.
Pattern Recognit., Sep. 1999, pp. 1033–1038.[9] O. Le
Meur, J. Gautier, and C. Guillemot, ―Examplar-based
inpainting based on local geometry,‖ in Proc. 18th IEEE Int.
Conf. Image Process., Sep. 2011, pp. 3401–3404.
[10] O. Le Meur and C. Guillemot, ―Super-resolution-based
inpainting,‖ in Proc. 12th Eur. Conf. Comput. Vis., 2012,
pp. 554–567.
[11] S. Dai, M. Han, W. Xu, Y. Wu, Y. Gong, and A.
Katsaggelos, ―SoftCuts: A soft edge smoothness prior for
color image super-resolution,‖ IEEE Trans. Image Process.,
vol. 18, no. 5, pp. 969–981, May 2009.
[12] W. T. Freeman, T. R. Jones, and E. C. Pasztor,
―Example-based superresolution,‖ IEEE Comput. Graph.
Appl., vol. 22, no. 2, pp. 56–65, Mar.–Apr. 2002.
[13] D. Glasner, S. Bagon, and M. Irani, ―Super-resolution
from a single image,‖ in Proc. IEEE Int. Conf. Comput. Vis.,
vol. 10. Oct. 2009, pp. 349–356.
[14] H. Chang, D.-Y. Yeung, and Y. Xiong, ―Super-
resolution through neighbour embedding,‖ in Proc. IEEE
Comput. Vis. Pattern Recognit., vol. 1. Jun.–Jul. 2004, pp.
275–282.
International Journal of Advanced Research in Computer Engineering & Technology (IJARCET)
Volume 3 Issue 11, November 2014
ISSN: 2278 – 1323 All Rights Reserved © 2014 IJARCET 3648
[15] Y. Wexler, E. Shechtman, and M. Irani, ―Space-time
video completion,‖ in Proc. IEEE Comput. Vis. Pattern
Recognit., Jun.–Jul. 2004, pp. I-120–I-127.
[16] Z. Xu and J. Sun, ―Image inpainting by patch
propagation using patch sparsity,‖ IEEE Trans. Image
Process., vol. 19, no. 5, pp. 1153–1165,
May 2010.
[17] S. Di Zenzo, ―A note on the gradient of a multi-image,‖
Comput. Vis., Graph., Image Process., vol. 33, no. 1, pp.
116–125, 1986.
[18] J. Weickert, ―Coherence-enhancing diffusion filtering,‖
Int. J. Comput. Vis., vol. 32, nos. 2–3, pp. 111–127, 1999.
[19] J. Kopf, W. Kienzle, S. Drucker, and S. B. Kang,
―Quality prediction for image completion,‖ ACM Trans.
Graph., vol. 31, no. 6, p. 131, 2012.
[20] A. Bugeau, M. Bertalmío, V. Caselles, and G. Sapiro,
―A comprehensive framework for image inpainting,‖ IEEE
Trans. Image Process., vol. 19, no. 10, pp. 2634–2644, Oct.
2010.
[21] D. D. Lee and H. S. Seung, ―Algorithms for non-
negative matrix factorization,‖ in Advances in Neural
Information Processing System. Cambridge, MA, USA:
MIT Press, 2000.
[22] A. Buades, B. Coll, and J. Morel, ―A non local
algorithm for image denoising,‖ in Proc. IEEE Comput. Vis.
Pattern Recognit., vol. 2. Jun. 2005, pp. 60–65.
[23] P. Pérez, M. Gangnet, and A. Blake, ―Poisson image
editing,‖ in Proc. SIGGGRAPH, 2003, pp. 313–318.
[24] N. Komodakis and G. Tziritas, ―Image completion
using efficient belief propagation via priority scheduling and
dynamic pruning,‖ IEEE Trans. Image Process., vol. 16, no.
11, pp. 2649–2661, Nov. 2007.
[25] Y. Boykov, O. Veksler, and R. Zabih, ―Efficient
approximate energy minimization via graph cuts,‖ IEEE
Trans. Pattern Anal. Mach. Intell., vol. 20, no. 12, pp.
1222–1239, Nov. 2001.
[26] Y. Boykov and V. Kolmogorov, ―An experimental
comparison of mincut/ max-flow algorithms for energy
minimization in vision,‖ IEEE Trans.
Pattern Anal. Mach. Intell., vol. 26, no. 9, pp. 1124–1137,
Sep. 2004.
[27] J. Yedidia, W. Freeman, and Y. Weiss, ―Constructing
free energy approximations and generalized belief
propagation algorithms,‖ IEEE Trans. Inf. Theory, vol. 51,
no. 7, pp. 2282–2312, Jul. 2005.
[28] A. Blake and A. Zisserman, Visual Reconstruction.
Cambridge, MA, USA: MIT Press, 1987.
[29] Y. Pritch, E. Kav-Venaki, and S. Peleg, ―Shift-map
image editing,‖ in Proc. Int. Conf. Comput. Vis., Sep. 2009,
pp. 151–158.
[30] K. He and J. Sun, ―Statistics of patch offsets for image
completion,‖ in Proc. 12th Eur. Conf. Comput. Vis., 2012,
pp. 16–29.
[31] J. Hays and A. A. Efros, ―Scene completion using
millions of photographs,‖ ACM Trans. Graph., vol. 26, no.
3, p. 4, 2007.
[32] O. Whyte, J. Sivic, and A. Zisserman, ―Get out of my
picture! Internetbased inpainting,‖ in Proc. 20th Brit. Mach.
Vis. Conf., 2009, pp. 1–11.
Breezy Joy A received the B.E degree in
Electronics and Communication
Engineering from Anna University,
Chennai, 2013.She is currently doing his
master of engineering in Communication
systems in Pet engineering college,
vallioor. Her areas of interests include
image processing, communication and
wireless systems.
Rekha C received the B.E degree in
Electronics and Communication Engineering
from Anna University, Chennai, 2005 and
M.E degree from Anna University,
Tirunelveli, 2011. She is currently working
as an Assistant Professor in Department of
Electronics and communication in Pet
engineering College, Vallioor. Her research
areas include antennas theory, microwave
systems,transmission lines and wave guides.