Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T.

Lopes

A Richardson-Lucy Algorithm Using a Varying Point Spread

Function Along the Iterations

Gevaldo L. de Almeida [email protected]

Instituto de Engenharia Nuclear / Reator Argonauta
Comissao Nacional de Energia Nuclear
Rio de Janeiro, 21941-972, Brazil

Maria Ines Silvani [email protected]

Instituto de Engenharia Nuclear / Reator Argonauta
Comissao Nacional de Energia Nuclear
Rio de Janeiro, 21941-972, Brazil

Erica S. Souza [email protected]

Instituto de Engenharia Nuclear / Reator Argonauta
Comissao Nacional de Energia Nuclear
Rio de Janeiro, 21941-972, Brazil

Ricardo T. Lopes [email protected]

Universidade Federal do Rio de Janeiro, COPPE, Centro de Tecnologia
Cidade Universitaria Bloco G, Ilha do Fundao, 21945-970 Rio de Janeiro - RJ, Brazil

Abstract

Image restorations with the Richardson-Lucy algorithm suffer the usual drawback imposed by the
constraint of a constant Point Spread Function – PSF as unfolding function. Indeed, even when
the image exhibits a constant spatial resolution over its whole surface, an important aspect is that
as the iterations advance, the overall resolution is improved while the PSF remains constant. This
work proposes an algorithm which restores images by the Richardson–Lucy (RL) algorithm, using
however, a varying PSF as the iterations proceed. For this purpose, the PSF width is reduced to
cope with the last-achieved image resolution and the next iteration would be carried out with it.
The process is repeated until the PSF does not change significantly. A main point in this
procedure is how to evaluate the PSF tied to the image resolution. In this work this is performed
on the grounds that the global contrast increases with the resolution improvement, for many gray
pixels migrate towards darker or brighter regions. Hence, deconvolving an image with a steadily
increasing PSF width, somewhere a maximum global contrast would be reached, corresponding
to the best PSF. Synthetic, as well as experimental images deconvolved with the proposed
technique, outperform the final quality of the same ones treated with the original Richardson-Lucy
algorithm for any number of iterations. The algorithm and ancillary procedures have been
embedded into an ad hoc written Fortran 90 program capable to generate synthetic images and
process them and the real ones.

Keywords: Image Restoration, Deconvolution, Richardson-Lucy, Varying PSF.

1. INTRODUCTION
Techniques to overcome poor image resolutions may be split into two categories: efforts to
produce the best possible image directly from the oven or/and a post-treatment. Both categories
demand hard conceiving efforts, but the last one is by far more numerable, for it can be
performed with a proper software, rather than sophisticated equipment. As for the second
category, the ultimate goal is to restore the primordial image – that which would be obtained in
the absence of degrading agents – through a post-acquisition treatment. In spite of the inherent
difficulties tied to this inverse problem – or perhaps just because that – a plethora of methods

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 40
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

have been developed, improved and employed to treat images. These methods could be roughly
classified into two branches: those ones requiring the previous knowledge of the Point Spread
Function – PSF and those named Blind Deconvolution [1], [2], where both PSF and final image
are simultaneously retrieved. It is well known, however, that this procedure lacks robustness and
may eventually return the same entered image and a zero–width PSF [3]. Among the methods
requiring a known PSF, the Richardson–Lucy [4], [5], henceforth called RL in this work, is a very
popular algorithm thanks to its robustness, granted convergence to the most likely image,
regardless of the image employed as initial guess, and a low-sensitivity to uncertainties in the
PSF.

The RL algorithm works with a constant PSF along all iterations for any region of the image.
However, an image could exhibit different spatial resolutions at different regions due to the
features of the equipment used to acquire it. In this work, spatial resolution and PSF or PSF width
are interchangeable terms, for all of them refer to the same Full width at Half Maximum (FWHM),
– henceforth also called width or w in this work – of a bell-shaped distribution. This distribution
may emerge as the response of an imaging system to a point-like source, and in this case, it is
referred to the PSF of the system.

Since there is a defined relationship between the lens focal distance and the gaps of the object
and its image to the lens, there is only a certain range – depth of field – where the image is
acceptably sharp. This means that the PSF should be rather depth–variant, but unlike the SPIM
florescence microcopy [6] it is not possible to illuminate a selected plane. An alternative technique
replaces this illumination by cropping the PSF directly from the image itself, provided that it
contains small and isolated enough structures [7].

An image may also exhibit a spatially-variant PSF – as some astronomical ones – a phenomenon
caused by the optical aberration of the telescopes, for which some methods [8], [9] have been
developed to deal with.

Hence, in order to apply a constant PSF one should choose a PSF-width which yields the best
overall image quality.

A hybrid method [10], using blind deconvolution and the RL algorithm has been as well
developed, applying a PSF, iteratively determined, but employed solely to recover synthetic
single-plane images of small sizes. Another work [11] determined the PSF by smoothing several
of them, obtained through the known technique of differentiating the edge response functions
(ERF). However, one of the main points addressed in the present work – namely the dealing with
a variable spatial resolution over the image pattern – was not addressed, for the ERFs were
acquired for a single plane. More recently, an evaluation of the PSF width based upon edge
detection and fuzzy logics [12] has been proposed for restoration of the images of satellites. It
however demands spatially invariant, and weakly or moderately degraded images.

Hence a best PSF – which would yield the best overall image quality – should be somehow
employed. However, even if the best PSF were known, there is still another constraint: along the
deconvolution, the image is continuously being improved but the unfolding PSF is kept constant
as required by the RL algorithm, i.e., the image is deconvolved with a shifted PSF width.

This mismatch is eliminated in this work by reducing the PSF width as the iterations proceed in
order to cope with the resolution of the last deconvolved image. Such a procedure requires the
knowledge of the resolution of the image so far achieved. This is performed with an algorithm
based upon the concept of Global Contrast (G–value for short) [13] as addressed in section 2.

For the same number of iterations, the proposed algorithm demands a greater computational
effort. Yet, the higher quality of the final restored image outperforms by far that obtained with the
conventional RL approach for any number of iterations, making thus worthwhile the additional
computational effort. Furthermore, the lower number of iterations required to reach the steady-

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 41
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

state precludes an excessive noise amplification, a shortcoming which led to the development of
damped versions of the RL algorithm [14], [15] to mitigate this phenomenon. Another plague
infesting the RL algorithm is the artifact appearing due to the discontinuity at the image
boundaries, as illustrated in the section 3.1, which led to the development of techniques to
mitigate it, such as [16].

2. METHODOLOGY
The proposed methodology is not tied to any type of acquisition system, as it uses solely the
information provided by the image itself. Synthetic images have been also generated with an ad
hoc written program to evaluate the soundness, capabilities and limitations of the methodology.
This approach allows a free choice of relevant parameters otherwise unfeasible or too expensive
to achieve.

2.1 Deconvolution with a Varying PSF Along the Iterations

The flow diagram sketched in Figure 1 highlights the differences between the conventional RL
procedure and the proposed one.

FIGURE 1: Conventional and novel procedures to restore images by the Richardson-Lucy algorithm.
st
Dashed lines represent actions performed only at the 1 iteration. Unlike the original procedure, the novel
one uses a varying PSF which is determined prior to each iteration.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 42
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

In the first one, the Original Image, a specified constant PSF, and an initial Image Guess are
delivered to the RL algorithm, which using always the original image and the last deconvolved
one performs a sequence of iterations with the same PSF until a certain condition is reached.
Unlike the PSF and the Original Image which are used for all iterations, the Image Guess is used
only once – a feature represented by dashed lines – and can be the original image itself, as used
in this work, or any one of the same size and same pixel/mm.

In the novel procedure, the iterations are carried out in a similar way, except that the New PSF
arising from the last deconvolved image is used to deconvolve the next one. Its assessment –
PSF Determination – is performed through the concept of Global Contrast as addressed in
section 2.1.1. The Original Image is used only once, for – as the Image Guess – it is replaced by
the last deconvolved one.

2.1.1 Determination of the Best PSF

An essential requirement to vary the PSF along the iterations is to determine its value after the
last iteration in order to apply it in the next one. A PSF is fully characterized by its shape and
width w, but it is assumed in this work that it has a Gaussian distribution. An image, however,
may exhibit different resolutions over its surface obliging thus the researcher to choose an
effective one, i.e., that yielding the best overall image quality. For that, one would have to scan a
broad range of PSF widths, and for each of them to carry out about 20-30 iterations. The final
images (one for each PSF width) should then be compared.

A visual inspection, besides its subjectiveness, would be a hard, boring and cumbersome task to
make the comparison. Instead, this work employs an algorithm which transduces the Global
Contrast [13] into a single number. This is done by searching for the PSF width w yielding the
best image (highest G-value) after 3 iterations performed with the conventional RL approach. It is
still necessary to scan a broad range of PSF widths but only 3 iterations (instead of 20-30) for
each of them. After an initial PSF width w, a chosen increment is added to w and a new set of 3
iterations is carried out. Further details can be found elsewhere [13].

The G–value increases with w, reaches a maximum, and then decreases again. Solely 3
iterations are sufficient to reach an acceptable steady value, making this approach an affordable
technique. The w–value yielding this maximum is assigned as the best PSF width and used for
the image deconvolution by the conventional RL algorithm. For the varying PSF algorithm, the
same approach is applied, but the best PSF width is determined after the 3 iterations carried out
for each entered w. The G–value is determined as follows:

1
M N M N 
G   u (i, j ). (i, j ). u (i, j ).1   (i, j )  (1)
 i1 j 1   i1 j 1 
 (i, j )  0 u (i, j )  um
for (2)
 (i, j )  1 u (i, j )  um
for (3)
M N
um  M .N  . u (i, j )
1
(4)
i 1 j 1

where:

u (i, j ) = Pixel value at the point (i, j).

M, N = No. of columns and lines of the image matrix.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 43
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

As stated by equation (4), um is the average pixel–value of the whole image, which is used to
classify all pixels into two classes: below and above it. The sums of all pixel–values occurring
above and below um constitute the numerator and denominator of equation (1) respectively,
defining the G-value, while the bump function i, j defined by equation (2) and equation (3)
classifies the pixel–values as below and above um respectively.

This formalism is justified as follows: after a deconvolution with a given PSF width w, many pixels
in the gray region shift towards darker or brighter regions due to the penumbra reduction. Hence,
the histogram tails raise with a consequent lowering of the remaining profile, for its area must be
conserved.

A higher fraction of pixels at darker and brighter zones means a higher contrast arising from a
better resolution. Therefore, the sum of all pixel–values above um would increase while that below
it would decrease, resulting into a higher G–value. This is illustrated with the histogram in Figure
2 arising from raw and deconvolved gamma–ray radiographic images of an old-fashioned film-
camera. Further details can be found elsewhere [13]. Yet, for the sake of completeness,
additional information concerning the determination of the best PSF-width, including some data
from [13] are addressed in the next paragraphs.

FIGURE 2: Actual histograms of a raw and a deconvolved image produced by gamma-ray radiography [13].
The resolution improvement shifts some pixels from gray zones toward brighter and darker ones raising the
tails and lowering the middle zone.

As observed in Figure 3, reproduced from [13], 3 iterations carried out with the conventional
Richardson-Lucy algorithm with a constant PSF-width are sufficient to achieve an acceptable
convergence. Both graphs have been obtained as follows.

An initial PSF-width is used as a Deconvolution w. After a certain number of iterations, varying

from 1 to 5, the global contrast is computed for the input image. Next, an increment is added to
that initial w and a new global contrast is determined. It initially increases with w, reaches a
maximum and then decreases. The abscissa of this maximum defines the best w for the
deconvolution with the Richardson-Lucy algorithm.

It can be noticed that for the synthetic image (left), the convergence to the best w occurs already
at the 1st iteration for each input w. The experimental image (right) – a thermal neutron

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 44
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

radiograph of an aeronautical turbine stator blade – requires however 3 iterations (for each input
w, as well). This difference is most likely caused by the different object thicknesses. Indeed, while
the synthetic image emulates the neutron radiograph of a flat object, the experimental one arises
from a thicker device, which casts a broader range of spatial resolutions.

The best PSF-width yields always the best overall image quality, for a given number of iterations.
If this number increases, that quality becomes improved, but the best PSF-width remains the
same. Indeed, the Fig. 3 shows an increasing global contrast with the number of iterations but the
peak epicenter keeps its position. Hence, 3 iterations used to determine the best w, constitute
simply a representative sample. Should this number be replaced by a larger one, no different
result would arise, but the CPU time would be enlarged with no advantage at all.

The best w determined for the original image, may be then employed as a constant PSF-width by
the deconvolution with the conventional Richardson-Lucy algorithm. For the same procedure
using a varying PSF-width, as here proposed, this parameter must be determined prior to each
deconvolution.

Therefore, an image restoration performed with a varying PSF involving n deconvolutions, would
require 3.s.n operations, where s is the number of PSF widths used to scan a given domain
encompassing the best w. The best w for the original image lies in the domain between zero and
a certain unknown maximum, which should be conservatively set as a large value. For the
subsequent images, that maximum is assigned as the best w obtained for the previous image,
making thus possible to define a domain. A further processing can be then automatically
performed by the program as addressed in section 2.4.

FIGURE 3: Convergence to the best PSF-width w, as posited by the Global Contrast approach. A thermal
neutron radiograph of a turbine stator blade (right) requires a larger number of iterations than the synthetic
image (left), most likely due to the broader range of spatial resolutions [13].

2.2 Synthetic Images

Although not developed in this work, the global contrast algorithm deserves a special analysis
due to its paramount importance in the proposed Richardson-Lucy algorithm using a varying PSF.
This analysis encompasses the algorithm response to several image patterns, e.g., different
spatial resolutions and contrasts, as expected to occur in general imaging. To achieve this task,
synthetic images are generated, making it possible to produce a customized image pattern.

This section therefore addresses 3 main purposes:

a) Generate images with a wide range of contrasts and spatial resolutions.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 45
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

b) Verify the soundness of the global contrast algorithm through its response to a wide
range of contrast and spatial resolutions as addressed in section 2.2.1.
c) Apply the proposed Richardson-Lucy with a varying PSF algorithm to the synthetic
images to verify the consistency of the obtained results (Section 3.1).

Once the topics a) and b) are properly addressed and shown their soundnesses, then, it is
possible to rely on the results arising at topic c).

Optical images may exhibit regions with different spatial resolutions and contrasts. In this work,
they are produced by an ad hoc program simulating a geometrical arrangement comprised by a
test-object, a surface light source and a screen where the image is projected. The maximal
achieved brightness is normalized to a pixel intensity of 65,535.

The test-object is constituted by 25 thin square blades of different transparencies, attached to a

fully transparent support, as sketched in Figure 4.

FIGURE 4: Test-object (left) and the arrangement to generate the image (right). Each pixel on the projection
screen – one is highlighted – is hit by the photons from the source and eventually attenuated by a blade.
Their contributions are ruled by the parameters L, D, g and by the blade opacity.

They cast individual shadows of different brightnesses and blurs, depending upon their
transparencies and gaps to the projection screen respectively. The parameters L, D and g rule
the spatial resolution of the generated images. A resolution of 0.05 mm/pixel has been assigned
to the image projected on the screen.

The algorithm to generate the synthetic images deals solely with 3D analytical geometry and the
Lambert’s law of attenuation. A projection screen receives the photons coming from a DxD
square source of light at a distance L, after passing through the object. For simplicity, only one
attenuating blade contained in the plane at a gap g from the projection screen is represented. It
can be realized that the L/D ratio rules the maximal angle between a photon path z(k,l,m,n) and
the normal to the source element where it comes from.

It is assumed that the photons are isotropically emitted by the source elements, and do not
undergo any dispersion or reflection but solely attenuation as they traversed the square blades.
The highlighted pixel on the projection screen is within a penumbra region because it receives all
photons coming from the highlighted source element at the bottom, but only part of them –

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 46
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

depending upon the attenuation factor of the intercepting square blade – coming from its upper
companion. This penumbra region grows with D/L and g.

To obtain the intensity map of the image cast on the screen of M x N pixels, a light source of K x
L square elements is placed at a distance L from the screen. Each pixel value p(m,n) depends
upon the angle k,l,m,ndefined by the coordinates on the screen and on the light source as
expressed by equation (5).

This value depends upon the source intensity, its distance to the screen and the attenuation of
th
the i square blade. All these parameters are condensed into a single value normalized to the
65,535 limit of a tiff-image.

K L
p(m, n)   exp f i .(k , l , m, n). I 0 .L / z k , l , m, n  (5)
k 1 l 1

where:

p(m,n) = Pixel value on the detector: m=1 to M, n=1 to N.

fi = Light attenuation factor of the blade i. i=1 to 25.
k,l,m,nBump function: 1 if the photon hits the blade i, 0 otherwise.
zk,l.m,nLength of the straight line connecting the pixel (m,n) on the screen with the (k,l)
element on the light source.
I0 = Reference intensity.
L = Distance source-screen.
K = No. of elements along the source width.
L = No. of elements along the source height.
M = No. of pixels along the image width on the screen.
N = No. of pixels along the image height on the screen.

The bump function value is assessed through the interception of the straight line z(k,l,m,n) with
the blade i. The reference intensity I0 encloses the source intensity. Any number can be chosen
for I0 since the matrix p(m,n) will be normalized to the maximum tiff-value of 65,535. The last term
L/z(k,l,m,n) is a geometric correction to cope with the photons arriving at the screen in a non-
perpendicular fashion. The M x N matrix is then stored for further processing and plotting by the
same Fortran program embedding the described algorithm.

2.2.1 Response of the Global Contrast Algorithm to Different Image Configurations

The deconvolution with a varying PSF requires its determination prior to each iteration. Due to the
paramount importance of this determination to the proposed RL algorithm, some typical cases are
addressed. Images may exhibit a huge number of different patterns, precluding thus a
comprehensive evaluation of the algorithm capability and robustness to deal with all of them. Yet,
it is possible to analyze its response to some specific cases by manufacturing virtual test-objects
at customer’s design. As this task requires solely 3D analytical geometry and program writing, it is
a matter of trade-off how far one should go. Within this frame, some typical test-object
configurations are addressed in the next sections, each of them yielding a different image pattern.
All of them are comprised of 25 attenuating blades after a chess pattern, distributed along several
gaps to the screen. The images cast by them on the screen have been inverted for the sake of a
better visualization.

2.2.1.1 Response to an Image Cast by a Single Plane Object

An object comprised of a single plane parallel to the screen – as that described in section 2.2, but
compressed to a zero thickness – represents an instance with a variable contrast and a constant
spatial resolution. It simulates the conventional photo of a chess-pattern flat square where the
single elements exhibit different brightnesses. Since their gaps to the camera are the same, their

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 47
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

blurs in the image would as well be identical. The resulting degraded image, its deconvolved
companion (50 iterations, w=4.2 pixels) and its undegraded version (the flat object attached
directly to the screen) are shown in Figure 5. The best PSF to restore the degraded image with
the conventional RL algorithm has been determined by the G-value approach, yielding a width of
4.2 pixels as shown in the same figure. The test-object size has been limited to 34x34 mm to
spare processing time.

FIGURE 5: Degraded image and its restoration by the original RL algorithm. The attenuating blades were at
15-mm gap to the screen. The best PSF width for the deconvolution was furnished by the graph. An
undegraded image is also shown for reference.

2.2.1.2 Response to Images Cast by Multi-plane Objects at Reversed Configurations

This section analyzes the impact of the brightness of the square blades and their gaps to the
screen upon the determination of the best w. It would be equivalent in a conventional photo, as if
the bright subjects where farther from the camera than the darker ones, and another photo were
o
taken with the inverse configuration. To perform this simulation, a rotation of the object at 180
around its vertical axis, would invert the gaps of the high and low–attenuating square blades to
the screen. Each blade contributes with its own gap to the best w, but the brighter the shadow it
casts on the screen, the greater its contribution. Therefore, low-attenuating blades – exhibiting a
corresponding low-contrast – become so faint when placed at large gaps that can be
overwhelmed by the prevailing background. Under such a condition, the high-attenuating blades,
closer to the screen, drift the best w towards their own positions.

Mutatis mutandis, the high-attenuating blades despite their larger gaps still exert a great influence
thanks to their high brightnesses. Indeed, as shown in Figure 6, when the high-attenuating blades
(casting bright shadows, for the images have been inverted) are located at large gaps, the
corresponding curve L exhibits a maximum G value at w=4.5 pixels. In the inverse configuration –

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 48
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

curve S – the G value reaches its maximum at w=1.8 pixels, i.e., the overall image quality should
be better. This forecasting is corroborated by the attached images.

It can be noticed that the G values themselves reach different amplitudes. This happens because
the image generated by blades at large gaps is more luminous than its companion as inferred
from the attached images. Indeed, for high-attenuating blades at short gaps, the low-attenuating
ones would be located at large gaps, and thus, could not contribute significantly to the overall
image luminosity due to the overwhelming prevailing background.

FIGURE 6: High-attenuating blades at short gaps (curve S), displaces the best w to a value (1.8 pixels) tied
to the average of their positions, while the images of the farther low-attenuating blades become so faint that
are overwhelmed by the background. In the reversed configuration, the high-attenuating blades still exert its
influence thanks to the high brightness of their cast images which shift the best PSF to w=4.5 pixels.

2.2.1.3 Response to Images Cast by Multi-plane Objects of Different Thicknesses

This section addresses the response of the Global Contrast algorithm to images exhibiting the
same contrast range as in previous section, but different spatial resolution ranges. To generate
them, 5 test-objects of several thicknesses – containing also 25 square blades under a chess
pattern – have been employed. A minimum gap from the closest square blade to the screen has
been kept at 5 mm for all objects, while the farthest one has been placed at 10, 15, 25, 35 and 45
mm. The intermediate blades have been regularly distributed between these limits.

To analyze the impact of these different image patterns upon the best w, the global contrast
algorithm has been applied. As shown in Figure 7, a larger range of the blade-screen gap yields a
larger best w. As expected, farther blades to the screen cast on it images with poorer resolutions,
and thus, a larger w.

Five different normalized profiles (G x PSF width w) are shown, one for each test-objects
thickness: 5, 10, 20, 30 and 40 mm, encompassing the blade–screen gap ranges 5-10, 5-15, 5-
25, 5-35 and 5-45 mm respectively. The profile broadening with the object thickness should
somewhat be expected, since the individual PSF widths w, related to each blade shadow are
more spread out. At any rate, these results match with those obtained in previous sections.

One can observe that the profile for the 5-10 mm gap at the left of the graph, unlike its smooth
companion 5-45 mm, exhibits a strong ripple. It most likely comes from arithmetic truncation and
roundness, since the written Fortran program deals with discrete pixel intensities. Yet, as this
drawback occurs at low PSF widths, it would affect only the last iterations, where a steady state
should already be reached.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 49
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

FIGURE 7: Response of the algorithm to 5 different gaps ranging from 5-10 to 5-45 mm. The closest insert
is always 5 mm apart from the screen, while the others are equally distributed until the maximum gap. As
expected, large gaps increase the best w.

As inferred from the analysis performed along this section 2.2.1, different image configurations –
of resolution and contrast – yield different PSF widths and should be accordingly deconvolved,
after the concept of best PSF. Hence, even if the conventional RL algorithm applies this PSF at
st
the 1 iteration, it is unquestionable that the resulting deconvolved image is improved. Therefore,
the PSF that should be applied to deconvolve this new image should have a narrower width w, in
order to match with the new achieved resolution. Nevertheless, the soundness of this concept
must be verified through deconvolution of synthetic and experimental images with the
conventional RL algorithm, as well as with its modified version employing a varying PSF as here
proposed. This topic will be addressed in section 3.

2.3 Experimentally Acquired Images

The technique to restore images by the RL algorithm with a varying PSF is proposed on the
grounds that the image undergoing deconvolution is being improved along the iterations. Hence,
each image obtained after a given iteration has its own overall resolution and should be
accordingly deconvolved, i.e., with a certain best PSF width w, which yields the best quality after
each iteration.

Once this value is obtained for the original image – as determined by the global contrast
algorithm – it can be used as the unfolding constant PSF width in the usual RL algorithm or
applied to the modified one. At any rate, the values of w do not depend upon the equipment or
means employed to acquire the image, but solely on the information contained in its pattern.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 50
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

Within this frame, besides the synthetic images, two experimentally acquired neutron and
gamma-ray radiographic images have been selected as examples, due to their wide range of
contrasts and resolutions. Details of their acquisition are superfluous and beyond the scope of
this work.

2.4 Automatic Deconvolution with a Varying PSF Along the Iterations

As the best w is defined by the maximum G–value, the process is automatized avoiding thus a
cumbersome task to the researcher. After the technique employed in this work, the PSF–width
domain is scanned with a chosen subinterval in order to detect the maximum G–value. The
abscissa of this maximum on the graph G x w defines the best PSF width to deconvolve the
image. As shown in Figure 8, each image i is deconvolved with its own best PSF width wi
previously determined by the global contrast algorithm.

For this determination, the RL and the global contrast algorithms are applied to the image i. A
conventional RL deconvolution with 3 iterations for each PSF width along the domain is then
carried out. The resulting global contrast is plotted against the PSF width i, until a maximum is
reached. This width is assigned as the Best PSF width, which it is input into the same image i
together with the conventional RL algorithm which performs one single iteration, yielding the
th
image i+1, and the procedure is repeated until the Best PSF width for the i image does not
change significantly.

Except for the original IMAGE 0, where a maximum PSF width could not a priori be assigned, for
all the subsequent ones this maximum is obviously the PSF width of the last image. This limit
reduces the domain to be scanned, saving thus CPU time. Yet, a certain tolerance is allowed to
deal with expected ripples in the curve Best PSF x Iteration Number as will be seen in section 3.

FIGURE 8: Automatic procedure to apply the Richardson Lucy with a varying PSF along the iterations. Each
iteration is carried out with the RL algorithm and a PSF width wi previously extracted from the image itself
with the Global Contrast algorithm. The iterative process is interrupted when the PSF width wi stabilizes.

As the iterations proceed, the spatial resolution, i.e., the PSF-width (best w) to deconvolve the
current image decreases, diminishing thus the domain encompassing it. A tolerance for the
higher limit of this domain is assigned, avoiding thus an eventual impact of ripples. This limit is
defined as Lr=f.w, where f is a parameter greater than 1. (in this work a conservative value of 1.1
has been chosen) and w is the spatial resolution from the previous image.

A shorter subinterval is required to scan a narrower domain, in order to assure a proper number

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 51
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

of points on the graph global contrast x PSF-width. The program then, uses a subinterval s=h.w
where h is a constant parameter previously chosen by the customer.

3. RESULTS
3.1 Synthetic Images
A synthetic image, as described in section 2.2, has been used as benchmark to check the
soundness of the developed algorithm. The left graph in Figure 9 depicts a family of normalized G
x w curves from images deconvolved with the proposed procedure in the range 0–7 iterations.
So, 4.5 pixels is the PSF width for the iteration 0, i.e., the original image.

Prior to each iteration, the best w is determined through the maximum G–value arising from
equations (1–4). The right graph simply reproduces the related maxima versus the iteration
number for the sake of a better visualization. For both algorithms (conventional or modified)
subsequently deconvolved images exhibit a decreasing w. Yet, for the varying PSF option, the
best PSF width drops faster demonstrating that it is more effective than the conventional one
using a constant PSF. Indeed, the upper curve in the right graph – expressing the best w
obtained along the iteration under this condition – exhibits higher PSF widths, except obviously
st
for the 1 iteration. Due to the asymptotic behavior, the iterations have been interrupted at the
th
11 one. Both options, however, exhibit artifacts and a trend to shrink the squares as shown in
Figure 10.

As this phenomenon is not tied to the proposed technique – but as well known, to the RL
algorithm itself – it does not constitute a drawback to its implementation and application. At any
rate, better images are achieved with a varying PSF. An undegraded image is also shown for
reference.

FIGURE 9: Family of curves G-value x PSF width (left) and their related best w x iteration number (right) for
deconvolution with a varying PSF (lower curve). Deconvolution with a constant PSF (upper curve) is less
efficient as the best w drops at a slower rate and reaches a higher value at the asymptotic region.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 52
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

FIGURE 10: Synthetic images unfolded with constant and varying PSF widths, ratify the higher performance
of the last approach. Artifacts arise with both procedures but the unfolding with a varying PSF yields a better
image quality. An undegraded image is also shown for reference.

3.2 Experimental Images

Experimental neutron and gamma-ray radiographic images of an old-fashioned film-camera have
been elected due their large range of contrasts and resolutions. The means, conditions and
parameters for their acquisition will not be addressed, as they are irrelevant for the image
treatment. Both images undergo the same treatment but are separately dealt, for the sake of
organization.

3.2.1 Thermal Neutron Radiographic Image

As in the synthetic case, each image is deconvolved with its best w determined through the G–
value. The related graphs and images are shown in Figure 11 and Figure 12 respectively. It can
be noticed that 10 iterations carried out with a varying PSF outperform those processed with a
constant value, even for 50 iterations. The first option, however, in spite of the lower number of
iterations, demands a higher computational effort. Yet, as the whole process is managed in an
automatic fashion by the program itself, the only drawback is the long required CPU time.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 53
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

FIGURE 11: Family of curves G-value x PSF width W (left) and the related Best w x Iteration No. (right).
th
Iterations were interrupted at the 11 one, for the best w converges to about 0.5 pixel.

FIGURE 12: Original and deconvolved neutron radiographic images of a film-camera. 10 iterations carried
out with a variable PSF outperform those done with a constant value even for 50 iterations. The first option
however, despite the lower number of iterations, demands a higher computational effort.

3.2.2 Gamma-Ray Radiographic Image

The results obtained with a gamma-ray radiographic image are presented in Figure 13. It can be
noticed that the original image exhibits a much better overall resolution than that achieved with
neutrons, an outcome tied to nuclear physics parameters which will not be addressed here. As
for the restored images, those deconvolved with a varying PSF also exhibit a better quality.
Indeed, 8 iterations carried out with a varying PSF outperform 50 iterations with a constant PSF.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 54
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

FIGURE 13: Original and deconvolved gamma-ray radiographic images of a film-camera. Eight iterations
undertaken with a varying PSF outperform those done with a constant value, even for 50 iterations.

3.3 Cross-Check of the Varying PSF Approach

As inferred from the previously presented images and graphs, a varying PSF along the iterations
yields better images than a constant one. However, due to this unique RL approach, it is highly
advisable to perform an additional cross-check of the procedure.

It has been already shown that an image deconvolved with a constant PSF width – as determined
by the global contrast algorithm – arising from the image itself, exhibits a poorer quality than that
deconvolved with a varying PSF. It is therefore important to verify what would be the result, if the
image were deconvolved (using a constant PSF) with the narrower PSF width emerged at the
end of a process employing the varying PSF technique. An important point of this verification is to
ratify that only a varying PSF along the iterations produces the best results, i.e., the initial or the
final resolution (arising from the varying PSF option) do not produce the best image when applied
to the original RL algorithm.

The impact of the PSF width of 6.5 pixels onto the neutron radiographic image deconvolved with
this constant value is presented in Figure 14. When this value is used as initial PSF width in the
varying PSF approach, after 10 iterations, the image exhibits a PSF width of 0.65 pixels, as
previously shown in the graphs of Figure 11. A deconvolution by the original RL algorithm with
this constant value of 0.65 pixels – i.e., the final resolution achieved at the end of the 10 iterations
with a varying PSF– likewise does not yield the best image as shown in Figure 14, an outcome
which is ratified by the graph in Figure 15. Indeed, the ratio 2.7/0.65 indicates quantitatively that
(for this image) the proposed modified RL algorithm produced a final image about 4 times better
than those ones arising from the original one.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 55
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

FIGURE 14: Neutron radiographic image deconvolved with a constant PSF of 0.65 pixels (the resolution
achieved at the end 11 iterations with a varying PSF). The original image and that one deconvolved with a
varying w are also shown for comparison.

FIGURE 15: Global Contrast x PSF-width for the neutron radiographic image deconvolved with 10 iterations
under a constant PSF-width of 0.65 pixels, i.e., the value reached by the same radiograph deconvolved with
10 iterations under a varying PSF, starting with 6.5 pixels.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 56
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

3.4 Comparison with Other Works

To the best of the authors knowledge, no work dealing with a RL algorithm using a varying PSF
along the iterations has been yet reported in the literature. Therefore, a comparison can be done
only with other different techniques. Among the immense number of developed techniques, blind
deconvolution is more close related to this work, for it as well, determines the PSF, though
simultaneously with the final deconvolved image.

The Figure 16 shows images of a spiral CT temporal bone (a), restored with the blind
deconvolution approach (b), as given by [2]. The original image (a) has been as well restored in
the present work with the conventional (c) and the proposed (d) RL algorithms respectively. Both
images, (a) and (b), have been retrieved from the original paper as a png file, amplified 2x using
the bilinear interpolation option of ImageJ, which converted it to a jpg format and delivers a matrix
with pixel values. The jpg-values of this matrix have been then converted to tiff format, precluding
thus, that the integer pixels values arising from the convolution of the PSF with the image matrix,
could be too early zeroed.

FIGURE 16: Original spiral CT image of a temporal bone (a), and its blind-deconvolved companion (b) [2].
Deconvolved images of the raw image (a) with the Richardson-Lucy algorithm under a constant PSF(c) and
with a varying PSF along the iterations (d) are shown for comparison.

As expected, the image processed with a varying PSF exhibits a better quality than its companion
processed with a constant one. Yet, a visual inspection shows that its quality is inferior to the
bind-deconvoluted image. It is therefore worthwhile to compare this visual evaluation with the
quantitative outcome arising from the global contrast technique as follows.

After the graph Global Contrast x w shown in Figure 17 (left), the blind-deconvolved image (b),
exhibits a spatial resolution of 0.45 pixels. This outcome from an image acquired by an
independent researcher, corroborates once more, that the global contrast approach – an
essential tool supporting the proposed varying PSF approach – behaves consistently, i.e., its G x
W curve exhibits a maximum.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 57
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

The curve Best w x Iteration No. (right), for the deconvolution of the same original image (a)
under a varying PSF, exhibits its typical behavior. An included zoom of its asymptotic tail shows
the Best w crossing the determined PSF-width for the blind-deconvoluted image (0.45pixels) -
th
represented by the horizontal line – at the 11 iteration and fairly stabilizing afterwards at circa of
0.3 pixels, an outcome that does not match with the visual evaluation. Although not possible to
assert that this is caused by the compulsory digging and digital treatment – imposed by the
unavailability of original image (a) – it is as well not possible to exclude this explanation.

Until other studies - with adequate images - could be conducted by independent researchers, it is
not possible to analyze the performance of the proposed method with regard to other techniques.
Nevertheless, the examples here presented demonstrated the superior performance of applying a
varying PSF along the iterations to the R-L algorithm, but further verifications should be carried
out by independent researchers.

FIGURE 17: Global Contrast x w profile arising from the global contrast approach (left), for the blind-
deconvolved spiral CT image of a temporal bone [2]. The same raw image deconvolved with the varying
PSF technique yields a final best w at 0.28 pixels as observed at the asymptotic region of the curve Best w x
Iteration No.(right).

Whether the abscissa of this maximum really represents the best w to unfold any image at all, is a
matter of verification through a visual inspection with many different images.

Yet, as far as the images here analyzed are concerned, both the varying PSF along the iterations
and its ancillary global contrast tool are sound procedures.

4. DISCUSSION
The posited technique to restore images by the RL algorithm with a varying PSF-width along the
iterations, yielded consistent results, for synthetic, as well as for actual experimentally acquired
images. Once verified that, besides the results obtained in Ref [13], the global contrast algorithm
reacted consistently to several other different image configurations, a solid basis for its
application could be asserted. Yet, it should be pointed out that any other approach capable to
replace the visual evaluation by a single number (as done by the global contrast technique) would
be applicable to the proposed method.

The basic idea behind the method is to unfold an image with its own spatial resolution or PSF-
width. If two researchers receive the same images to restore, but the image delivered to one of
them had been undergone a previous partial deconvolution, (with 1 or 2 iterations), they certainly
would realize that their images exhibit different qualities, and would infer that they arise from
systems possessing different resolutions.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 58
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

The final image furnished by a system is the outcome of a convolution between the primordial
(unknown) image and the PSF of the system. With a fully characterized PSF the image would be
retrieved simply by solving a linear system of equations. Unfortunately, this ideal solution may be
applied only – as a demonstration – to extreme minuscule (a couple of pixels) matrices, and thus,
useless for the imagery field. Indeed, for useful images (even those of modest sizes), the huge
2 2
coefficient matrices (a nxn image requires a n xn coefficient matrix), roundness and truncation
required to cope with the required integer pixel values for the image, generated ill-posed
matrices, precluding to solve the system. This constraint led to the development of iterative
techniques, as the RL algorithm.

But even with a fully characterized PSF, the primordial image could never be retrieved, due to this
ill-posed inverse problem. Only the most likely resembling image could be achieved. Hence, if it is
already a hard task to get an approximate image, even with an exact PSF, it is not an efficient
and reasonable approach to abdicate of it in favor of a less precise value. In other words, the
st
PSF-width employed at the 1 iteration – disregarding whether it is correct or not – should not be
used in the remaining ones, since the deconvolved images are being improved, and their PSF-
width concomitantly narrowed.

Furthermore, the varying PSF approach is expected to be more forgiving to uncertainties in the
st
initial PSF with because – disregarding whether the 1 deconvolved image has been improved or
additionally degraded due to an improper w – in the subsequent iterations the deviation would be
corrected. This correction should occur because the global contrast algorithm would determine
st
the PSF width of the 1 deconvolved image (improved or degraded), and its value employed to
correctly deconvolve it.

Switching back to the example of the two researchers, in order to perform a proper deconvolution
of their images with the RL algorithm, they need a fully characterized PSF, i.e., its shape and
width (FWHM). If they use an eventual technique to evaluate the PSF, certainly they would find
different outcomes, because the partially deconvolved image would exhibit a better quality due to
the improvement performed by those previous iterations. So, if the 1st researcher keeps its
evaluated PSF-width, along the whole deconvolution procedure (as prescribed by conventional
the Richardson-Lucy algorithm), it would be carried out under a shifted PSF-width. Indeed, while
nd st
the 2 researcher started with a narrower value (due to the previous iterations), the 1 one would
nd rd
still be applying its broader PSF at the 2 , 3 and further iterations.

An image deconvolved with a zero PSF-width would not be modified at all. It would be equivalent
to an acquisition performed with an ideal system. Under such a circumstance, the system would
have not degraded the primordial image, and thus, there would be nothing to be corrected.
Indeed, it is a well-known blind deconvolution drawback, that this approach sometimes delivers
an unmodified input image together with a zero PSF-width. On the other hand, a PSF with a
larger width than that tied to the image itself, not only would not improve it properly, but could
even degrade it. Somewhere, between these limits lies the best PSF-width, reproducing the
effective spatial resolution of the image.

As the PSF shape is very difficult to obtain, it is usually replaced by a radially symmetric 3D-
function. In this work a Gaussian has been employed with fairy good results. However, it is not
possible to assure that it is the best function to restore the images presented in this work (nor
other images as well). It is therefore an open field for researchers working in image restoration, to
seek for better unfolding functions (PSF shapes) and algorithms to replace the image pattern by a
single number representing its overall quality, as done by the global contrast approach.

An important feature of an algorithm is naturally its robustness. However, it only can be assured
through a large number of restorations with different images by different researchers. Such a task
should be performed through many researchers, due to the huge number of runs involved, and to
assure independent verifications. The impact of the proposed technique in the field of imagery

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 59
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

constitutes, as well, an even wider – and hard – field to explore, due to the true zoo of image
restoration methods. A comprehensive comparative study would be a herculean burden.

This work aims rather a far less ambitious target, namely, to propose a procedure to vary the PSF
width along the iterations in the RL algorithm, and compare the resulting images with those
arising from a constant PSF.

Future research aiming at to strengthen this technique and establish its robustness, capability
and limits, could include, e.g., the study of various PSF shapes and its impact on the final image
quality, the search of other approaches to evaluate the PSF, the combination with other
techniques in hybrid methods, such as, stopping criteria for the iterations, damping RL algorithms,
elimination or reduction of artifacts, and naturally a robustness evaluation of the proposed
method, through an extensive verification involving images of several spatial resolution an
contrast ranges.

5. CONCLUSION
A novel procedure is proposed to restore images by the Richardson–Lucy algorithm, using
however a varying Point Spread Function – PSF along the iterations. Within this approach, the
spatial resolution of the image obtained at the last iteration is used as the PSF width (w) for the
next iteration, until it does not change significantly. The w-value is assigned as that yielding the
greatest Global Contrast, on the grounds that it increases with the image spatial resolution, for
many gray pixels migrate to darker or brighter regions. Although applied in this work only to
synthetic, neutron and gamma-ray radiographic images – as well as, to a spiral CT image for the
sake of comparison – the procedure may in principle be applied to images of any kind, since all
the required data are provided by the images themselves. The only alien required parameter is
the PSF shape. A 3-D Gaussian function has been used in this work, but other ones could
produce better results, an eventuality which deserves a more comprehensive evaluation involving
different functions, and images. As the spatial resolutions of the images are essential to perform
this evaluation, they should be somehow determined by using the technique employed in this
work, namely the global contrast concept, or any other suitable one. At any rate, despite the
unusual use of the Richardson-Lucy algorithm, the promising results obtained in this work
deserve a deeper analysis from researchers working in the image restoration field.

6. REFERENCES
[1] D. Kundur and D. Hatzinakos, “Blind Image deconvolution”, Signal Processing Magazine
IEEE, Vol. 13, No. 3, pp. 43-64, 1996.

[2] M. Jiang, G. Wang, “Development of Blind Image Deconvolution and its Applications”,
Journal of X-Ray Science and Technology Vol.11, pp. 13-19. 2003

[3] F. Tsumuraya, N. Miura, N. Baba, “Iterative Blind Deconvolution Method using Lucy`s
Algorithm”, Astronomy and Astrophysics, Vol. 282, pp. 699-708, 1994.

[4] W. H. Richardson, “Bayesian-based Iterative Method of Image Restoration”, Journal. of the

Optical Society of America, Vol. 62, No. 1, pp. 55-59, 1972.

[5] B. L. Lucy, “An Iterative Technique for the Rectification of Observed Distributions”, The
Astronomical Journal, Vol. 79, No. 6, pp.745-754, 1974.

[6] M. Temerinac–Ott, O. Ronneberger, R. Nitschke, W. Driever, H. Burkhardt, “Spatially-variant

Lucy-Richardson Deconvolution for Multiview Fusion of Microscopical 3D Images”, in IEEE
International Symposium on Biomedical Imaging, Chicago, IL, USA, pp. 899-904, 2011.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 60
Gevaldo L. de Almeida, Maria Ines Silvani, Erica S. Souza & Ricardo T. Lopes

[7] J. B. de Monvel, E. Scarfone, S. Le Calvez, M. Ulfendahl, “Image-Adaptative Deconvolution

for Three-Dimensional Deep Biological Imaging”, Biophysical Journal, Vol. 85, pp.3991-
4001, 2003.

[8] M. Faisal, A. D. Lanterman, D. L. Snyder, R. L. White, “Implementation of a Modified

Richardson-Lucy Method for Image Restoration on a Massively Parallel Computer to
compensate for Space-variant Point Spread of a Charge-coupled-device Camera”, Journal
of the Optical Society of America A, Vol. 12, Issue 12, pp. 2593-2603, 1995.

[9] H. Cheong, E. Chae, E. Lee, G. Jo, J. Paik, “Fast Image Restoration for Spatially Varying
Defocus Blur of Imaging Sensor”, Sensors, Vol. 15, No.1, pp. 880-898, 2015.

[10 D. A. Fish, A. M. Brinicombe, E. R. Pike, J. G. Walker, “Blind deconvolution by Means of

the Richardson-Lucy algorithm”, J. Opt. Soc. America A, Vol. 12, No. 1, pp. 58-65, 2005.

[11] F. Krejci, J. Jakubek, J. Dammer, D. Vavrik, “Enhancement of Spatial Resolution of

Roentgenographic Methods”, Nuclear Instruments and Methods in Physics Research A ,
Vol. 607, pp. 208-211, 2009.

[12] F. Aouinti. M. Nasri, M. Moussaoui, “Estimating the Size of the Point Spread Function using
Fuzzy Logic for Blind Restoration of Satellite Image”, International Journal of Recent
Contributions from Engineering , Science & IT - iJES, Vol.5, No. 4, pp. 48-59, 2017.

[13] G. L. Almeida, M. I. Silvani, “A Novel Algorithm for Blind Deconvolution applied to the
Improvement of Radiographic Images”, in AIP Conference Proceedings, Sao Sebastiao,
SP, Brazil, Vol. 1529, pp. 95-99, 2013.

[14] R. L. White, “Image Restoration using the Damped Richardson – Lucy Method”,
The Restoration of HST Images and Spectra II Space 01 Telescope Science Institute, R. J.
Hanisch and R. L. White, Eds, 1994.

[15] B. Masschaele, M. Dierick, L. V. Hoorebeke, P. Jacobs, J. Vlassenbroeck, V.Cnudde,

“Neutron CT Enhancement by Iterative de-blurring of Neutron Transmission Images”,
Nuclear Instruments and Methods in Physics Research A, Vol. 542, Issues 1-3, pp. 361-
366, 2005.

[16] M. Bertero, P. Boccacci, “A Simple Method for the Reduction of Boundary Effects in the
Richardson-Lucy Approach to Image Deconvolution”, Astronomy and Astrophysics, Vol.
437, No. pp. 1-6, 2005.

International Journal of Imaging Processing (IJIP), Volume (13) : Issue (4) : 2019 61