Fast nas rif algorithm using iterative conjugate gradient method

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.2, April 2014
DOI : 10.5121/sipij.2014.5206 63
FAST NAS-RIF ALGORITHM USING ITERATIVE
CONJUGATE GRADIENT METHOD
A.M.Raid1
, W.M.Khedr2
, M. A. El-dosuky3
and Mona Aoud1
1
Mansoura University, Faculty of Computer Science and Information System
2
Zagazig University, Faculty of Science
3
Mansoura University, Faculty of Computer Science and Information System
ABSTRACT
Many improvements on image enhancemen have been achieved by The Non-negativity And Support
constraints Recursive Inverse Filtering (NAS-RIF) algorithm. The Deterministic constraints such as non
negativity, known finite support, and existence of blur invariant edges are given for the true image. NAS-
RIF algorithms iterative and simultaneously estimate the pixels of the true image and the Point Spread
Function (PSF) based on conjugate gradients method. NAS-RIF algorithm doesn’t assume parametric
models for either the image or the blur, so we update the parameters of conjugate gradient method and the
objective function for improving the minimization of the cost function and the time for execution. We
propose a different version of linear and nonlinear conjugate gradient methods to obtain the better results
of image restoration with high PSNR.
KEYWORDS
blurred image, NAS-RIF algorithm, Image restoration, Point Spread Function (PSF), Conjugate Gradient
(CG) and Peak Signal to Noise Ratio (PSNR).
1. INTRODUCTION
In image restoration field, A prior information such as the PSF of the imaging system and the
observation noise [1]. When the blur effect is yielded on the objects, the PSF and the observation
noise are unknown a priori. The identification and the restoration of the blurred images is a
challenging problem in the world. The algorithm based on NAS-RIF [2,3] is proposed in order to
identify and restore the images.
The scene which consists of a finite support object against a uniformly black, grey, or white
background has been restored by The NAS-RIF technique. The restoration algorithm of NAS-RIF
includes Recursive filtering of the blurred image to reduce a convex cost function.
The different versions of linear and nonlinear conjugate gradient methods are used for
minimization of the NAS-RIF cost function. The conjugate gradient depends on gradient of an
image at determine parameters. We update the CG’s parameter and compare the results to

64
improve the restored image and the PSNR. The experimental results and comparison with the CG
parameters show that the restoration effect is improved obviously.
The rest of this paper organize in three sections, The next section introduces fundamental of
image restoration, section three explains NAS-RIF algorithm in more details with the update of
CG parameters and section four shows the results of experimental.
2. HISTORICAL REVIEW
2.1 Image Restoration
Most of images can be modeled by using the standard degradation model as the following:
݃ሺ‫,ݔ‬ ‫ݕ‬ሻ = ℎሺ‫,ݔ‬ ‫ݕ‬ሻ ∗ ݂ሺ‫,ݔ‬ ‫ݕ‬ሻ + ߟሺ‫,ݔ‬ ‫ݕ‬ሻ ሺ1ሻ
Where,
1- ሺ‫,ݔ‬ ‫ݕ‬ሻ discrete pixel coordinate of the image
2- ݃ሺ‫,ݔ‬ ‫ݕ‬ሻ image which has been blurred
3- ݂ሺ‫,ݔ‬ ‫ݕ‬ሻ original image
4- ℎሺ‫,ݔ‬ ‫ݕ‬ሻ the point spread function (PSF)
5- ߟሺ‫,ݔ‬ ‫ݕ‬ሻ additive noise
6- 2‫ܦ‬ convolution operator
The above model shows that the degraded image݃ሺ‫,ݔ‬ ‫ݕ‬ሻ, the original image݂ሺ‫,ݔ‬ ‫ݕ‬ሻ, and the noise
ߟሺ‫,ݔ‬ ‫ݕ‬ሻ, are linearly combined and as a result the linear image restoration problem is the
problem of retrieving the image from degraded image.
The form of bandwidth reduction of an ideal image owing to the imperfect image formation
process is known as blurring. Relative motion between the camera and the original scene, or an
optical system that is out of focus is caused this problem. Atmospheric turbulence, aberrations in
the optical system and relative motion between the camera and the ground are caused blurs when
aerial photographs are produced for remote sensing purposes. The optical images are not the only
example of blurring, for example spherical aberrations of the electron lenses corrupted the
electron micrographs, and X-ray scatter corrupted the CT scans suffer.

65
In addition to these blurring effects, any recorded image has been corrupted by noise. The
medium may introduce the noise through which the image is created (random absorption or
scatter effects), by the recording medium (sensor noise), by measurement errors because of the
limited accuracy of the recording system, and by quantization of the data for digital storage.
The reconstruction or estimation of the uncorrupted image from a blurred and noisy one is
concerned by the field of image restoration which also known as image deconvolution or image
deblurring. Essentially, the inverse of the imperfections in the image formation system operation
is performed on the image by the image restoration. A priori information are known about the
characteristics of the degrading system and the noise in the use of image restoration methods. In
practical situations, however, one may not be able to obtain this information directly from the
image formation process. blur identification aims at estimating the attributes of the imperfect
imaging system from the observed degraded image itself prior to the restoration process. The
blind image deconvolution is combination of image restoration and blurs identification [4].
The purpose of the blind image deconvolution[5] is to obtain a reliable and effective estimate for
the original image from the degraded version. To achieve this task, we use the partial information
about the imaging process as a reference to deconvolve the true image and the PSF from the
blurred image.
Existing blind deconvolution[6] methods include:
1) A priori blur identification methods.
2) Zero sheet separation methods.
3) ARMA parameter estimation methods.
4) Nonparametric methods.
A priori blur identification methods perform the blind deconvolution by identifying the PSF prior
to the restoration. Typically, the PSF to be of a known parametric form is assumed by these
methods. The unknown associated parameters must determine before restoration. For certain
PSFs, the frequency-domain zeros of the PSFs determine the parameters uniquely. Once the zeros
of Hሺu, vሻ are identified, the parameters associated with hሺx, yሻ can be estimated accordingly.
The a priori blur identification methods are relatively simple to achieve and require low
computational complexity. But they have three major drawbacks. (1) They require the knowledge
of the form of the PSF. (2) some PSFs (such as Gaussian PSF) simply don’t have frequency
zeros. (3) Additive noise can mask the frequency-domain nulls of and then degrade the
performance. The lack of sufficient 2D polynomial algebra is one of the difficulty problems in
the blind deconvolution.
The zero sheet separation algorithm [7] is such a factorization scheme. The valuable insight into
the blind deconvolution problems in multiple dimensions is provided by the method of zero sheet
separation. However, it suffers from several major problems. The First problemof this method is
highly sensitive to noise. The Second problem, the method is prone to inaccuracy for larger
images.
The true image as an autoregressive (AR) process and the PSF as a moving average (MA)
process are modeled by Blind deconvolution using ARMA parameters estimation. The blurred
image is therefore modeled as the autoregressive moving average (ARMA) process. The
advantage of the ARMA parameter estimation methods is that they are less sensitive to noise
because the models already take into account the noise. The major problem of these methods has
the risk of ill-convergence to local minima. In addition, the total number of parameters cannot be

66
very large for practical computations. Usually, the restoration by these methods is not uniquely
decided unless additional assumptions are made about the PSFs. They are therefore referred to as
the nonparametric methods. Belongs to this category, there exist many algorithms including:
1) Iterative blind deconvolution (IBD) method.
2) Simulated Annealing (SA) method.
3) Nonnegativity and support constraints recursive inverse filtering (NAS-RIF).
In this paper, we consider the following details about the imaging process, the true image, and the
PSF:
1- The linear model of the degradation as shown in Figure 1.
2- The grey, black, or white background of the image is uniformly.
3- A priori such as the nonnegative of the true image and its support; the smallest
rectangle encompassing the object of interest known as the support (the region of
support shown in Figure 2).
4- The inverse of the PSF exists, and both the PSF and its inverse are absolutely
summable.
5- When the image background is black, the sum of all PSF pixels is assumed to be
positive, which occurs in almost all image processing applications.
2.2 NAS-RIF Algorithm
The assumption of non negativity on the true image is marked by NAS-RIF algorithm [8] that is,
the algorithm supposes the true image pixel intensity values will be nonnegative. The knowledge
of a known support is as well assumed. The only assumptions made on the PSF however, is that it
is absolutely summable and that it has an inverse ℎିଵ
ሺ‫,ݔ‬ ‫ݕ‬ሻ which is also absolutely summable.
This method has an advantage is that it does not require the PSF to be of known finite extent, as
do the other methods; this information is often difficult to get. The NAS-RIF algorithm is shown
in Figure 3. It consists of a variable FIR Filter uሺx, yሻ with the blurred image gሺx, yሻ as input. The
estimation of the true image ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ. Represents the output of this filter this estimate is passed
through a nonlinear filter, which uses a non-expansive mapping to project the estimated image
into the space representing the known characteristics of the true image. The error signal to update

67
the variable represents filter uሺx, yሻ the difference between the projected image ݂ே௅ሺ‫,ݔ‬ ‫ݕ‬ሻ and
݂ሺ‫,ݔ‬ ‫ݕ‬ሻ.
If the image assumes to be nonnegative with known support, the NL block of Figure 3 shows the
projection of the estimated image onto the set of images that are nonnegative with given finite
support. Then, the negative pixel values within the region of support must be zero, and the
background grey-level, ‫ܮ‬஻ ݅‫ݏ‬ the pixel values outside the region of support. The cost function ‫ܬ‬
used in the restoration procedure is defined as:
‫ܬ‬ = ෍ ൣ݂መே௅ሺ‫,ݔ‬ ‫ݕ‬ሻ − ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ൧
ଶ
∀ሺ௫,௬ሻ
ሺ2ሻ
Where,
݂መே௅ሺ‫,ݔ‬ ‫ݕ‬ሻ = ൝
‫ܮ‬஻
0
݂መሺ‫,ݔ‬ ‫ݕ‬ሻ
(3)
Where ‫ܦ‬௦௨௣ suppose to be the set of all pixels inside the region of support, and ‫ܦ‬ഥ௦௨௣ suppose to
be the set of all pixels outside the region of support. Replace fመ୒୐ in equation (2) leads to
‫ܬ‬ = ෍ ൣ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ − ‫ܮ‬஻൧
ଶ
ሺ௫,௬ሻ∈ ஽ഥೞೠ೛
+ ෍ ݂መଶሺ‫,ݔ‬ ‫ݕ‬ሻ ቈ
1 − ‫݊݃ݏ‬ሺ ݂መሺ‫,ݔ‬ ‫ݕ‬ሻሻ
2
቉
ሺ௫,௬ሻ∈ ஽ೞೠ೛
ሺ4ሻ
Where the ‫݊݃ݏ‬ሺ. ሻ defined as
‫݊݃ݏ‬ሺ݂ሻ = ൜
−1 ݂݅ ݂ < 0
1 ݂݅ ݂ ≥ 0
ሺ5ሻ
For globally minimizes ‫,ܬ‬ The parameter ‫ݑ‬ሺ‫,ݔ‬ ‫ݕ‬ሻ set to 0 for all ሺ‫,ݔ‬ ‫ݕ‬ሻ. This results to all black
solution in which a restored image ݂መሺ‫;ݔ‬ ‫ݕ‬ሻ = 0 for allሺ‫,ݔ‬ ‫ݕ‬ሻ. To prevent this trivial solution, the
sum of all the PSF coefficients assumed to be positive,i.e.
‫݉ݑݏ‬ ሺ ℎሺ‫,ݔ‬ ‫ݕ‬ሻሻ > 0 ሺ6ሻ

68
Also the filtering doesn't cause any loss of total image intensity because the sum of inverse filter
coefficients is supposed to be 1, the only method for constraining the parameters is normalized
‫ݑ‬ሺ‫,ݔ‬ ‫ݕ‬ሻ at every iteration to fulfill this condition. Another method is add a third term to the cost
function which known as a penalty method. The overall function is then represented by
‫ܬ‬ = ෍ ݂መଶሺ‫,ݔ‬ ‫ݕ‬ሻ ቈ
1 − ‫݊݃ݏ‬ሺ ݂መሺ‫,ݔ‬ ‫ݕ‬ሻሻ
2
቉
ሺ௫,௬ሻ∈ ஽ೞೠ೛
+ ෍ ൣ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ − ‫ܮ‬஻൧
ଶ
+ ߛ ቎ ෍ ‫ݑ‬ሺ‫,ݔ‬ ‫ݕ‬ሻ
∀ሺ௫,௬ሻ
− 1 ቏
ଶ
ሺ7ሻ
The cost function consists of three components. The negative pixels of the image estimate inside
the region of support has been affected by The first component, the pixels of the image estimate
outside the region of support that are not equal to the background color have been affected by the
second component. The first component inhibits the pixels of the intermediate restorations from
becoming highly negative. The detail algorithm is shows in Algorithm (1).
Algorithm (1): NAS-RIF Algorithm
I. Definition:
• ݂௞ሺ‫,ݔ‬ ‫ݕ‬ሻ: ݁‫݁ݐܽ݉݅ݐݏ‬ ‫݂݋‬ ‫݁ݑݎݐ‬ ݅݉ܽ݃݁ ܽ‫ݐ‬ ݇‫ݐ‬ℎ ݅‫݊݋݅ݐܽݎ݁ݐ‬
• ‫ݑ‬௞ሺ‫,ݔ‬ ‫ݕ‬ሻ: ‫ܴܫܨ‬ ݂݈݅‫ݎ݁ݐ‬ ‫ݏݎ݁ݐ݁݉ܽݎܽ݌‬ ‫݂݋‬ ݀݅݉݁݊‫݊݋݅ݏ‬ ܰ௫௨ × ܰ௬௨ܽ‫ݐ‬ ݅‫݊݋݅ݐܽݎ݁ݐ‬ ݇
• ‫ܬ‬൫‫ݑ‬௞൯: ܿ‫ݐݏ݋‬ ݂‫݊݋݅ݐܿ݊ݑ‬ ܽ‫ݐ‬ ‫ݎ݁ݐ݁݉ܽݎܽ݌‬ ‫݃݊݅ݐݐ݁ݏ‬ ‫ݑ‬௞
• ∆‫ܬ‬൫‫ݑ‬௞൯: ݃‫ݐ݊݁݅݀ܽݎ‬ ‫݂݋‬ ‫ܬ‬ ܽ‫ݐ‬ ‫ݑ‬௞
II. Set initial condition (k=0):
• Set FIR filter ‫ݑ‬௞ሺ‫,ݔ‬ ‫ݕ‬ሻ to all zeros
III. At iteration (k): k=0, 1,2, ……
1) ݂መ௞ሺ‫,ݔ‬ ‫ݕ‬ሻ = ‫ݑ‬௞ሺ‫,ݔ‬ ‫ݕ‬ሻ ∗ ݃ሺ‫,ݔ‬ ‫ݕ‬ሻ
2) ݂መே௅ሺ‫,ݔ‬ ‫ݕ‬ሻ = ܰ‫ܮ‬ൣ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ൧
3) Minimization Routine to update FIR filter parameters.(conjugate gradient routine)
a) ൣ∆‫ܬ‬ሺ‫ݑ‬௞ሻ൧
்
= ൤
డ௃ሺ௨ೖሻ
డ௨ሺଵ,ଵሻ
డ௃ሺ௨ೖሻ
డ௨ሺଵ,ଶሻ
… ..
డ௃ሺ௨ೖሻ
డ௨൫ேೣೠ,ே೤ೠ൯
൨
Where
߲‫ܬ‬൫‫ݑ‬௞൯
߲‫ݑ‬ሺ݅, ݆ሻ
=
2 ෍ ݂መଶሺ‫,ݔ‬ ‫ݕ‬ሻ ቈ
1 − ‫݊݃ݏ‬ሺ ݂መሺ‫,ݔ‬ ‫ݕ‬ሻሻ
2
቉ ݃ሺ‫ݔ‬ − ݅ + 1, ‫ݕ‬ − ݆ + 1ሻ
ሺ௫,௬ሻ∈ ஽ೞೠ೛
+ 2 ෍ ൣ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ − ‫ܮ‬஻൧
݃ሺ‫ݔ‬ − ݅ + 1, ‫ݕ‬ − ݆ + 1ሻ
+ 2ߛ ቎ ෍ ‫ݑ‬ሺ‫,ݔ‬ ‫ݕ‬ሻ
∀ሺ௫,௬ሻ
− 1 ቏
b) ߚ = ൫< ∇‫ܬ‬൫‫ݑ‬௞൯ − ∆‫ܬ‬൫‫ݑ‬௞ିଵ൯, ∇‫ܬ‬൫‫ݑ‬௞൯ >൯/ ൫< ∇‫ܬ‬൫‫ݑ‬௞ିଵ൯, ∇‫ܬ‬൫‫ݑ‬௞ିଵ൯ >൯
c) ݂݅ ݇ = 0, ݀௞ = − ∆‫ܬ‬൫‫ݑ‬௞൯
‫ݐ݋‬ℎ݁‫,݁ݏ݅ݓݎ‬ ݀௞ = − ∆‫ܬ‬൫‫ݑ‬௞൯ + ߚ ݀௞ିଵ
d) ‫ݑ‬௞ିଵ = ‫ݑ‬௞ + ‫݀ݐ‬௞
4) k=k+1
5) goto Step 3

69
We update the parameters of conjugate gradient [9] function and the value of constant ߛ to be:
ߚ௞ = ൫< ∇‫ܬ‬൫‫ݑ‬௞൯, ∇‫ܬ‬൫‫ݑ‬௞൯ − ∆‫ܬ‬൫‫ݑ‬௞ିଵ൯ >൯/ ൫< ∇‫ܬ‬൫‫ݑ‬௞൯,∇‫ܬ‬൫‫ݑ‬௞ିଵ൯ >൯
ߛ = max ሺߚ௞ሻ
The algorithm is shows in Algorithm (2).
Algorithm (2): Fast NAS-RIF Algorithm
I. Definition:
• ݂௞ሺ‫,ݔ‬ ‫ݕ‬ሻ: ݁‫݁ݐܽ݉݅ݐݏ‬ ‫݂݋‬ ‫݁ݑݎݐ‬ ݅݉ܽ݃݁ ܽ‫ݐ‬ ݇‫ݐ‬ℎ ݅‫݊݋݅ݐܽݎ݁ݐ‬
• ‫ݑ‬௞ሺ‫,ݔ‬ ‫ݕ‬ሻ: ‫ܴܫܨ‬ ݂݈݅‫ݎ݁ݐ‬ ‫ݏݎ݁ݐ݁݉ܽݎܽ݌‬ ‫݂݋‬ ݀݅݉݁݊‫݊݋݅ݏ‬ ܰ௫௨ × ܰ௬௨ܽ‫ݐ‬ ݅‫݊݋݅ݐܽݎ݁ݐ‬ ݇
• ‫ܬ‬൫‫ݑ‬௞൯: ܿ‫ݐݏ݋‬ ݂‫݊݋݅ݐܿ݊ݑ‬ ܽ‫ݐ‬ ‫ݎ݁ݐ݁݉ܽݎܽ݌‬ ‫݃݊݅ݐݐ݁ݏ‬ ‫ݑ‬௞
• ∆‫ܬ‬൫‫ݑ‬௞൯: ݃‫ݐ݊݁݅݀ܽݎ‬ ‫݂݋‬ ‫ܬ‬ ܽ‫ݐ‬ ‫ݑ‬௞
II. Set initial condition (k=0):
• Set FIR filter ‫ݑ‬௞ሺ‫,ݔ‬ ‫ݕ‬ሻ to all zeros
III. At iteration (k): k=0, 1,2, ……
1) ݂መ௞ሺ‫,ݔ‬ ‫ݕ‬ሻ = ‫ݑ‬௞ሺ‫,ݔ‬ ‫ݕ‬ሻ ∗ ݃ሺ‫,ݔ‬ ‫ݕ‬ሻ
2) ݂መே௅ሺ‫,ݔ‬ ‫ݕ‬ሻ = ܰ‫ܮ‬ൣ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ൧
3) Minimization Routine to update FIR filter parameters.(conjugate gradient routine)
a) ൣ∆‫ܬ‬ሺ‫ݑ‬௞ሻ൧
்
= ൤
డ௃ሺ௨ೖሻ
డ௨ሺଵ,ଵሻ
డ௃ሺ௨ೖሻ
డ௨ሺଵ,ଶሻ
… ..
డ௃ሺ௨ೖሻ
డ௨൫ேೣೠ,ே೤ೠ൯
൨
Where
߲‫ܬ‬൫‫ݑ‬௞൯
߲‫ݑ‬ሺ݅, ݆ሻ
=
෍ ݂መଶሺ‫,ݔ‬ ‫ݕ‬ሻ ቈ
1 − ‫݊݃ݏ‬ሺ ݂መሺ‫,ݔ‬ ‫ݕ‬ሻሻ
2
቉ ݃ሺ‫ݔ‬ − ݅ + 1, ‫ݕ‬ − ݆ + 1ሻ
ሺ௫,௬ሻ∈ ஽ೞೠ೛
+ ෍ ൣ݂መሺ‫,ݔ‬ ‫ݕ‬ሻ − ‫ܮ‬஻൧
݃ሺ‫ݔ‬ − ݅ + 1, ‫ݕ‬ − ݆ + 1ሻ
+ max ሺߚ௞ሻ ቎ ෍ ‫ݑ‬ሺ‫,ݔ‬ ‫ݕ‬ሻ
∀ሺ௫,௬ሻ
− 1 ቏
b) ߚ = ൫< ∇‫ܬ‬൫‫ݑ‬௞൯, ∇‫ܬ‬൫‫ݑ‬௞൯ − ∆‫ܬ‬൫‫ݑ‬௞ିଵ൯ >൯/ ൫< ∇‫ܬ‬൫‫ݑ‬௞൯, ∇‫ܬ‬൫‫ݑ‬௞ିଵ൯ >൯
c) ݂݅ ݇ = 0, ݀௞ = − ∆‫ܬ‬൫‫ݑ‬௞൯
‫ݐ݋‬ℎ݁‫,݁ݏ݅ݓݎ‬ ݀௞ = − ∆‫ܬ‬൫‫ݑ‬௞൯ + ߚ ݀௞ିଵ
d) ‫ݑ‬௞ିଵ = ‫ݑ‬௞ + ‫݀ݐ‬௞
4) k=k+1
5) goto Step 3
3. SIMULATED RESULTS
We implement the original and updated conjugate gradient parameter in the NAS-RIF algorithm,
the objective function and compare the results on different Images samples with motion blur
effect using length=10, angle=0.05, the experimental results show in the following figure in
which the image restoration improve obviously.

70
Example 1:
Figure 4.1: (a) Original Image (b) Motion blurred image with length=10, angle=0.05
Figure 4.2: (a) Restored Image using NAS-RIF Algorithm, PSNR= 61.7979 dB
(b) Restored Image using Fast NAS-RIF Algorithm, PSNR= 63.9548 dB
Example 2:

71
(b) Restored Image using Fast NAS-RIF Algorithm, PSNR= 66.6734 dB
Example 3:
(b) Restored Image using Fast NAS-RIF Algorithm, PSNR= 57.3027 Db

72
4. CONCLUSION
In this paper, the parameters of conjugate gradient method and the objective function are updated
to improve the minimization of the cost function and the time for execution, and finally, a priori
knowledge of the region of support of the object is changed to a simple classification procedure.
REFERENCES
[1] D. Kundur, D. Hatzinakos, “A Novel Blind Deconvolution Scheme for Image Restoration Using
Recursive Filtering” [J], IEEE Transaction on Signal Processing, 46(2), 375-389(1998).
[2] Shirpur, D.Dhule, D.Maharastra, “Blind Image Restoration Methods”, Department of Computer
Engineering, R.C.Patel Institute of Technology, India, April 21, 2012
[3] B.Majhi, “Improvements in Blind Image Restoration”, Department of Computer Science and
Engineering National Institute of Technology Rourkela - 769 008, India, 2010
[4] D.Kundur and D.Hatzinakos, “Blind Image Deconvolution”, IEEE Signal Processing Magazine, vol.
13 (3), pp. 43-64, May 1996.
[5] T.F.Chan, C.K.Wong, “Convergence of the alternating minimization algorithm for blind
deconvolution”, Linear Algebra and its Applications, Volume 316, Issues 1–3, 1 September 2000,
Pages 259-285, ISSN 0024-3795.
[6] W.Wang, J.J.Zheng, S.Chen, S.Xu, H.Zhou, “Two-stage blind deconvolution scheme using useful
priors”, International Journal for Light and Electron Optics, Volume 125, Issue 4, February 2014,
Pages 1503-1506, ISSN 0030-4026.
[7] P. Premaratne, C.C. Ko, “Retrieval of symmetrical image blur using zero sheets“, Department of
Electrical Engineering, The National University of Singapore , Singapore, Vision, Image and Signal
Processing, Volume 148, Issue 1, February 2001.
[8] T.Xavier,”Development of image restoration techniques”, National Institute of Technology Rourkela,
May 2007.
[9] H. Yabe and M. Takano, “Global convergence properties of nonlinear conjugate gradient methods
with modified secant condition”, Computer Optim. Appl., 28, (2004), pp. 203-225.

Fast nas rif algorithm using iterative conjugate gradient method

Recommended

More Related Content

What's hot (20)

Viewers also liked (19)

Similar to Fast nas rif algorithm using iterative conjugate gradient method (20)

Recently uploaded (20)

Fast nas rif algorithm using iterative conjugate gradient method