Mr. Ashok.K
Abstract
Image decomposition is an ill-posed problem usually
addressed in various applications of image processing such as
image denoising, enhancement etc.. In normal image
denoising processes, it is not clear how to decompose the
image into multiple semantic components and thereby
removing those components which correspond to undesirable
noise patterns. Here in this paper, we introduces an image
decomposition framework using Discrete Wavelet
Transform(DWT) filters. The system first identifies and
learns an over-complete dictionary from the high spatial
frequency parts of the input image for reconstruction
purposes. Then performs unsupervised clustering on the
observed dictionary atoms via method of affinity
propagation. Once the proposed system identifies the image
components which are similar, then the proposed framework
automatically removes the noise contaminated image
components directly from the input image. The performance
of this approach depends on the relative denoising techniques
used and on the number of iterations of the algorithm and in
most of the cases we require more than hundred iterations
which increase the efficiency of the system. In this paper, we
propose a new and improved method for increasing the
overall efficiency of the system. Introduction of DWT filter
into the existing system for images corrupted by Gaussian
noise or rain noise are very useful because of its ability to
capture the signal energy in few energy transformation
values. Our experiments show that the proposed system
yields improvements in PSNR and thereby increases the
sharpness of images mainly at the edges.
Index terms: denoising; sparse representation; image
decomposition; clustering.
I.
INTRODUCTION
k(Ik,
k)
k||2
(1)
www.ijaert.org
TABLE I
THE NOTATIONS AND DESCRIPTIONS OF THE SYMBOLS IN THIS PAPER
II.
(3)
www.ijaert.org
III.
www.ijaert.org
(a)
(b)
(c)
(a)
(b)
(c)
(d)
(e)
(d)
Input
ij)
ijs(di,dj)-
- 1|
(4)
ii)(
ij)
(f)
Figure 4: extracting IH from the input image I by using
different LPF methods.(a),(b),(c) represents the filtered
output corresponding to bilateral, KSVD and BM3D methods
respectively. (d),(e),(f) represents the corresponding I H=I-IL
for bilateral, KSVD and BM3D methods respectively.
www.ijaert.org
Ground Truth
(a)
Input
(a)
(c)
(d)
Figure 5: Example rain removal results. Input represents the noisy version of the ground truth image with rain streaks. The rain
removed output by methods of (a) MCA based(b) bilateral (c)KSVD and (d)BM3D
IV.
www.ijaert.org
Ground Truth
Input
(a)
(b)
(c)
(d)
Figure 6: Example showing image denoising results. Input represents the noised version of ground truth image (with Gaussian
noise). (a),(b),(c),(d),represents the denoised output produced by methods of bilateral, KSVD, SURELET and BM3D respectively.
Besides rain removal, our proposed method can be applied
for removing Gaussian noise from images. for this task, as in
the case of rain removal we first decomposes the image into
IH and IL using any of the filtering techniques. The example
for producing high frequency component is illustrated in
figure. After obtaining IH, the dictionary DH is learned and
then the HOG features corresponding to each atoms d i are
extracted. HOG has no role in describing Gaussian noise,
while the presence of Gaussian noise would result in
undistinguishable bin/attributes in HOG features. The noise
free dictionary atoms exhibits dominant bin/attribute in their
HOG features. Even though The standard deviations of the
noise is not given here, the undesirable image patterns
corresponding to Gaussian noise are still able to identify
using our decomposition and clustering Framework. After
removing such noise patterns, the remaining HF components
are added to the spatial low frequency components I L and
thus the input image can be reconstructed. The examples are
shown in figure.
V.
FILTER TYPE
OUR RESULTS
Bilateral
19.574
19.21
KSVD
19.720
19.43
BM3D
19.673
19.50
www.ijaert.org
OUR RESULTS
29.8674
25.93
Bilateral
KSVD
26.9139
26.82
BM3D
26.7229
26.25
SURE LET
27.2070
26.69
FILTER TYPE
THE Table III shows the PSNR values obtained using our
different denoising approaches. From the results it is obvious
that the obtained improved denoising results than the
standard denoising approaches existing today.
VI.
CONCLUSION
REFERENCES
[1] J.M. Fadili, J. L. Starck, J. Bobin, and Y.Moudden, Image
decomposition and separation using sparse representations: an
overview, Proc. IEEE, vol. 98, no. 6, pp. 983994, Jun. 2010
www.ijaert.org