This document discusses image deblurring using sparse domain selection. It begins with an introduction that discusses sources of blur and the need for deblurring. It then provides an overview of image deblurring basics including modeling blur with a point spread function. The main method presented is an adaptive sparse domain selection approach that learns image structures to better model patches. It provides experimental results showing improved peak signal to noise ratio and structural similarity index values compared to other methods. In conclusion, the adaptive sparse domain selection is shown to significantly improve sparse modeling and image restoration results.
3. Where does blur come from?
Optical blur: camera is out-of-focus
Motion blur: camera or object is moving.
Why do we need deblurring?
Visually annoying
Wrong target for compression
Bad for analysis
Numerous applications
5. Resent increases in screen resolution of
Cameras, Camcorders, Television has
increased the need for image pre-
processing.
The deblurring appears in a wide range
of application fields, Photography,
medical imaging, recorders .etc.,
The ill-posed problem can be represented
as y = Dx+ n
6. Input blurred image
Classify the blur content
into high, low and no blur
Convert into each individual
patches for iteration
Sparse
representation
Check
convergence
Output image, sparse
representation
yes
no
Flow Chart of minimization algorithm
6
7. 𝒙 = 𝐚𝐫𝐠 𝐦𝐢𝐧
𝒙
‖𝒚 − 𝑫𝒙‖ 𝟐
The optimized method to solve the ill posed
problem is
A Point Spread Function (PSF) is simply the
Fourier Transform of the blur filter.
Different kinds of blur can be modeled with
a PSF. e.g. linear, Gaussian, etc.
Many dictionary learning methods used in
learning various image structures.
The problem ca be solved by Augmented
Lagrangian Method.
Image Deblurring Deconvolution Basics
8. Image Regularization with Adaptive Sparse Domain Selection
The popular restoration method is used to approach in image
regularization computed from recovered method
The blur estimation is estimated by PSF used in blind
Deconvolution.
The matrix method used in image boundaries with M X N & M
X L vector formation.
Figure: Spectra of kernal regularization matrices.
2 1
3 1
3 2
0
0 0.
0
x x
x x h
x x
𝑥2 −𝑥1 0
𝑥3 0 −𝑥1
0 𝑥3 −𝑥2
ℎ = 0.
9. Approaches & Objectives
Image processing is growing field so can
survive our life.
Ex. Consumer electronics,
Image restoration/reconstruction much required for
growing world.
Ex. The world changing towards high definition
3D, HD, UHD etc.
10. the images are preserving with respect to
blurring of an images.
The pixel sharp is increased.
The pixels of image Gaussian noise is reduced.
Original
image
Gaussian
noise
Ill-posed
problem
Gaussian
filter
Deblurred
image
11. $ 2 2
|| || || ||x y Hx xy
Image Deblurring Basis
• computed with frequency as they turn convolution in
time to multiplication in frequency.
• The deconvolution is the reverse of convolution
which cannot be directly computed.
• To solve ill-posed problem we have a clam of
algorithm called as least square minimization
𝒚 = 𝑯 ∗ 𝒙 + 𝒏
12. Although image contents can very a lot from image to image
(micro structure of images)
Eg. Edges, line segments & other elementary features.
The ASDS assets sparse domain with norm.
The patch selection criterion is to exclude the smooth
patches from training & guarantee contain edge structures
are involved in dictionary learning.
Both adaptive regularization & non-local similarity
regularization into the ASDS based sparse representation.
The different kernel size are tested in proposed methods.
𝓁1
13. Image contents can very a lot from image to image,
The human visual system employs a sparse coding strategy to represent images.
As a Clustering based method to choose the classes.
Figure: Comparison of deblurred images
PSNR=36.81dB, SSIM=0.8926
PSNR=37.86dB, SSIM=0.9540
14. Comparison of deblurred images with o/p of ASDS-TD1
( a. PSNR=37.81dB, SSIM=0.9526,
b. PSNR=38.64dB, SSIM=0.9720,
c. PSNR=39.79dB, SSIM=0.9833)
a. Non local image
c. Reconstructed image
b. Noise image
15. We proposed a sparse representation method.
Based on image deblurring and super resolution
method
The ASDS improves significantly the effectiveness of
sparse modeling and consequently the results of
image restoration.
An iterated shrinkage algorithm was
proposed to implement the proposed ASDS algorithm
with AReg.
16. Reference
1. R. G. Keys, “Cubic convolution interpolation for digital image processing,”IEEE
Trans. Acoust., Speech, Signal Process., vol. 29, no. 6,pp. 1153–1160, Dec. 1981.
2. A. Buades, B. Coll, and J. M. Morel, “Nonlocal image and moviedenoising,” Int. J.
Comput. Vis., vol. 76, no. 2, pp. 123–139, 2008.
3. J. A. Tropp and S. J.Wright, “Computational methods for sparse solution of linear
inverse problems,” Proc. IEEE, vol. 98, no. 6, pp. 948–958,Jun. 2010.
4. Daubechies, M. Defriese, and C. DeMol, “An iterative thresholding algorithm for
linear inverseproblems with a sparsity constraint,” Commun. Pure Appl. Math.,
vol.57, pp.1413-1457, 2004.
5. M. Elad and M. Aharon, “Image denoising via sparse and redundant
representations over learned dictionaries,” IEEE Trans. Image Process., vol. 15, no.
12, pp. 3736-3745, Dec. 2006.
6. J. Mairal, G. Sapiro, and M. Elad, “Learning Multiscale Sparse Representations for
Image and Video Restoration,” SIAM Multiscale Modeling and Simulation, vol. 7,
no. 1, pages 214-241, April 2008.
7. R. Rubinstein, M. Zibulevsky, and M. Elad, “Double sparsity: Learning Sparse
Dictionaries for Sparse Signal Approximation,” IEEE Trans. Signal Processing, vol.
58, no. 3, pp. 1553-1564, March 2010.
8. F. Šroubek and J. Flusser, “Multichannel blind deconvolution of spatially
misaligned images,” IEEE Trans. ImageProcess., vol. 14, no. 7,pp. 874–883, Jul. 2005.