ISSN (ONLINE): 2454-9762

ISSN (PRINT): 2454-9762

Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017

A METHOD OF TARGET TRACKING USING

INVERSE SYNTHETIC APERTURE RADAR HIGH
FREQUENCY ANALYSIS TECHNIQUE

ASHWINI A1
1
P.G Scholar, Dept. of ECE, Ponjesly College of Engineering,
Aalamparai,Tamilnadu, India
[email protected]

Abstract— Image denoising is an important NHA. Finally quantitative evaluation is

image processing task both as a process performed on image sequences of different
itself and as a component in other processes. carrier ships and airplane targets with the
The process of tracking the objects in space elimination of the scatterrers present in the
is gaining popularity with the advance of images thus reaching its specified
technology. Non-harmonic analysis is a high destination region successfully.
resolution frequency analysis technique that
improves noise removal accuracy in which Keywords—Non-harmonic analysis;
a multiframe marked point process model Multiframe marked point process;
for automatic target structure extraction by scatterrers.
tracking in inverse synthetic aperture radar I. INTRODUCTION
image sequences and the process of tracking
The detection of the scatterrers in the
the images throughout its movement has
image and the elimination of it seem to be a
been determined. To deal with scatterer
tedious task nowadays .The method for
scintillations and high speckle noise in the
classifying a ship based on a 3D model
ISAR frames the resulting target sequence
extracted from an ISAR image sequence has
by an iterative optimization process, the
been dealt in the existing model of the target
observed image data and various prior
recognition. Traditionally, Inverse Synthetic
geometric interaction constraints between
Aperture Radar (ISAR) image frames are
target appearances in the consecutive frames
classified individually in an automatic target
is considered. Non uniform regions that
recognition system. When information from
occur due to segmentation are analyzed by
different image frames is combined, it is
an extended NHA method called Mask

All Rights Reserved © 2017 IJARMATE 64

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
usually in the context of time averaging Cut (ALG) ideas to exploit their
overlapping of the images takes place and complementary strengths. It consists of three
remains a tedious task to remove statistically main parts: model building, initialization,
independent noise fluctuations between and delineation. For the initialization
images. The sea state induced variability of (recognition) part, a pseudo strategy is
the ship target projections between frames, employed and the organs are segmented
however, also provides additional slice by slice via the OAAM (Oriented
information about the target that can be used Active Appearance method). The purpose of
to construct a 3D representation of the target initialization is to provide rough object
scatterer positions. This model gives the localization and shape constraints for a latter
maximum of two to three frames. Numerous GC method, which will produce refined
denoising methods have been developed in delineation. It is better to have a fast and
diverse fields of research. Katkovnik et al. in robust method than a slow and more
their review suggested a classification of accurate technique for initialization. Christo
denoising algorithms according to two Ananth et al. [9] proposed a system which
features: local/non-local and uses intermediate features of maximum
pointwise/multipoint algorithms[6]. Priyam overlap wavelet transform (IMOWT) as a
Chatterjee and Peyman Milanfar proposed a pre-processing step. The coefficients derived
patch-based, locally adaptive denoising from IMOWT are subjected to 2D histogram
method based on clustering the given noisy Grouping. This method is simple, fast and
image into regions of similar form of unsupervised. 2D histograms are used to
geometric structure. In order to effectively obtain Grouping of color image. This
perform such clustering, features of the Grouping output gives three segmentation
local weight functions is used. These maps which are fused together to get the
weights are exceedingly informative and final segmented output. This method
robust in conveying reliable local structural produces good segmentation results when
information about the image even in the compared to the direct application of 2D
presence of significant amounts of noise. Histogram Grouping. IMOWT is the
Next, each region or cluster is built which efficient transform in which a set of wavelet
may not be spatially contiguous by learning features of the same size of various levels of
a best basis describing the patches within resolutions and different local window sizes
that cluster using principal components for different levels are used. IMOWT is
analysis[15]. Christo Ananth et al. [5] efficient because of its time effectiveness,
proposed a system in which an automatic flexibility and translation invariance which
anatomy segmentation method is proposed are useful for good segmentation
which effectively combines the Active results.This simple detector using operators
Appearance Model, Live Wire and Graph of several widths to cope with different

All Rights Reserved © 2017 IJARMATE 65

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
signal-to-noise ratios in the image is taken result shown only shows that the estimated
for processing. Femto second laser pulses 3D models are, in some sense, consistent
were used to excite second harmonic waves between image sequences, but there is no
from collagen harvested from rat tail tendon proof that this is because it reliably extracts
and a reference nonlinear crystal. Second 3D position information. This could be
harmonic interference fringe signals were partially confirmed by checking whether a
detected and used for image construction. similar rate of comparison could be achieved
Because of the strong dependence of second by using the original 3D scatterer models as
harmonic generation on molecular and tissue the template models. It is also not certain
structures. as a frequency analysis method, that it is only the 3D information that is
non-harmonic analysis (NHA), which has producing the ship discrimination. This
high accuracy for detached frequency could be confirmed by comparing the 3D
components is used and is only slightly performance against that of a classifier
affected by the frame length[3]. The based on the range profiles of the ship, or
spectrum subtraction method is one of the based on time-averaged feature based
most common methods by which to remove classifiers.
noise from a spectrum. Inverse Synthetic
Aperture Radar (ISAR) image shows the
two dimensional reflectivity of a target by
synthesizing the signal reflected from a
target .Its range resolution is decided by the
bandwidth. The cross-range is obtained by
the Doppler frequency caused by the relative
rotational motion between the target and the
radar. Though the method seems to be
simple the noise that are present in the
images is not completely extracted that lead II.PROPOSED METHOD
to the loss of data. Investigation into the 3D
model extraction method is still in its Image formation
preliminary stages, and the evidence for the The ISAR images is segmented into
usefulness of the technique is still only foreground and background classes by a
suggestive, and not conclusive. The binary Markov random field model to
proposed method brings vast advantageous decrease the spurious effects of speckle
over the existing system as it implies the use noise. The goal is to obtain a binary label
of the multi target analysis and elimination map labels correspond to the foreground and
of the noise by means of the point process background classes respectively. To get an
model has been delivered. Specifically, the initial estimation of the target axis segment,

All Rights Reserved © 2017 IJARMATE 66

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
detect the axis line using the Hough more efficient ones considering the data and
transform of the foreground mask. This prior constraints in parallel.
model enables to characterize whole target
sequences instead of individual objects Masked Non Harmonic Analysis
through exploiting information from entity
The frequency resolution of DFT is
interactions. Following the classical
limited by the integer period. If the period of
Markovian approach, each target sample
the signal in the analysis window has a non-
may only affect objects in its neighboring
integer value, sidelobes are observed in the
frames directly. This property limits the
results. In an analysis of DFT, the frequency
number of interactions in the population and
resolution is chosen according to the change
results in a compact description of the global
in analysis window size. However, NHA
sequence, which can be analyzed efficiently.
estimates the Fourier coefficients by using
Interaction potentials are responsible for
sinusoidal in frequency analysis using DCT.
involving temporal information and prior
In general, the frequency resolution depends
geometric knowledge in the model. Image
on the window size.
denoising using domain transformation often
employs a block unit called a “patch.” The
smoothing effectiveness of denoising The clustering performance taking
increases with the patch size. The non into consideration depends on effect of the
masked analysis is a high resolution noise. The original signal can be distorted by
frequency analysis technique that removes the removal of the sidelobe through
the noise removal accuracy. The mask can thresholding. This is the obstacle to
be with horizontal and the vertical alignment denoising. NHA can suppress sidelobes
of the pixel. The sum of products of the using high-frequency resolution. Image
individual pixels are alligned using the denoising using domain transformation often
neighbouring pixels. The mask is placed employs a block unit called a “patch.” The
over individual pixels and the masking smoothing effectiveness of denoising
operation is performed. The edges of the increases with the patch size. However, the
image is determined accurately. The likelihood of the edge mixing with the
boundary can be detected without any of the patches increases when using large patches.
blurring effect as the lower pixel values are The edge is transformed into the sinc
removed over the high frequency resolution. function by DFT. When part of the error
The edges are detected clearly and also signal in the frequency domain is removed
accurately without any loss of the data. by thresholding, the restored image includes
Iterative refinement algorithm, which scans a ringing artifact. The performance of
in each step the whole sequence and denoising method using natural images
attempts to replace the actual objects with provoides the analysis of images using DFT
and DCT leads to sidelobes. If the analysis

All Rights Reserved © 2017 IJARMATE 67

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
window contains regions of different
textures adjacent to each other, the
corresponding signal in the analysis window
is non-stationary. Hence, the input signal in
the analysis window needs to be masked.
The masking area is obtained from image
segmentation results. In segmentation, the
noisy image is classified into two regions,
the edge region and the texture region.
Moreover, texture region is classified into
multiple regions using spatial similarity.

Target Axis Extraction Of Patch Based

Images

To minimize the influence of signal non- Fig 2 Workflow of proposed method

stationarity, it is necessary to exclude edges As scatterers may appear and
and different textures from the analysis disappear in ISAR images on the basis of
window. Most of the ATC/ATR systems shadowing effects or weak scattering
that are based on radar images make use of a mechanisms, the estimation of the target’s
two-step approach to solve the problem: size may be incorrectly performed. By
features are extracted from the radar image observing a target for a longer period of
and then they are fed to a classifier that time, scatterers typically appear and
decides based on comparing such features disappear from image frame to frame.
with those that were previously stored in a Therefore, by using the ISAR image
database. sequences, such shortcomings may be
overcome. Sequences of ISAR images were
used previously to improve both ISAR
image formation and target’s classification
and recognition.

III. PROBLEM DEFINITIONS AND

NOTATIONS
The input of the proposed algorithm is an n-
frame long sequence of 2-D ISAR data,
imaged in the range-Doppler domain, which
contains a single ship (or airplane) target.

All Rights Reserved © 2017 IJARMATE 68

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
Let it is denoted by S the joint pixel lattice methods rely on the fact that the image
of the images and by s ∈ S a single pixel. acquisition and therefore the system
The amplitude of pixel s in frame t ∈ {1, 2 .. parameters are suitably tailored to handle
…. n} is marked with ξ. As the observed certain type of targets. As an example, the
values may vary in a wide amplitude range range and Doppler extension of the imaged
for a more compact data representation of area should be larger than the range and
the images the input maps by taking the Doppler occupation of the target. As the
logarithm of the observed amplitudes, latter depends on the radar-target dynamics,
thereafter we apply linear scaling for this condition is not necessarily verified.
normalization. The zero amplitude values Nevertheless, certain types of target and
are replaced with a small positive constant radar platforms typically produce
to avoid the calculation of log 0. The images predictable or at least bounded dynamics
corresponding to the sample frames in the that can give direct information about the
normalized log-amplitude domain are radar parameter settings that ensure that
displayed in gray scale. Apart from such a condition is 0 30 60 90 120 150 180
visualization, the logarithmic image 210 240 pixel intensity: normalized
representation suits well the widely adopted log−amplitudes empirical likelihood
log-normal statistical models of the ISAR foreground background. Histogram of
target segmentation. This paper is to intensity values over a 18-frame long
measure relevant features of the objects, subsequence. Foreground and background
such as length or orientation, which provide regions are manually distinguished satisfied.
us information for target identification and This fact allows us to define foreground–
behavior analysis. The investigated ISAR background segmentation in a
images provide only very limited straightforward way.
information about the superstructures of the
targets, identify stable bright points in the Foreground–Background Segmentation
images, called permanent scatterers, which
In the first step, the ISAR images is
can be tracked over the frames of the
segmented into foreground and background
sequence .These characteristic features are
classes by a binary Markov random field
produced by stronger scatterer from the
model to decrease the spurious effects of
illuminated objects.
speckle noise. The goal is to obtain a binary
IV DATA PREPROCESSING IN A DIRECT label map labels correspond to the
APPROACH foreground and background classes,
respectively. Assuming that the ξ-amplitude
Data preprocessing is needed to prepare the values in both classes follow log-normal
data for subsequent analysis. It should be distributions that model the log-amplitude
pointed out that some of the following posterior probabilities by Gaussian densities.

All Rights Reserved © 2017 IJARMATE 69

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
To experimentally validate this model, frame followed by a pair of morphological
foreground masks onto sample images is closing and opening iterations, where the
done and investigated the log-amplitude threshold corresponds to the value of the
feature statistics in the foreground and integral histogram. The preliminary mask
background regions, respectively. This too coarse for object shape investigations;
shows the histograms of the two classes, as a however, it proved to be appropriate for
result of evaluating 18 frames selected from estimation of posterior probabilities, as the
a 245-frame-long ISAR sequence with estimated Gaussian parameters differed only
uniform time intervals, and observed that the slightly from the supervised estimation
Gaussian approximation is valid. To results. If pixel is in the four-neighborhood
estimate the Gaussian distribution of pixel are in the S lattice then optimal
parameters of the foreground and foreground mask is derived through
background classes, one can choose either a minimizing the following MRF energy
supervised or an unsupervised approach. function.
Initial Alignment and Line Segment
Supervised models need manual Estimation
foreground–background evaluation of a few To get an initial estimation of the
sample frames; however, these key frames target axis segment, detect the axis line
have to be carefully chosen as the range of using the Hough transform of the foreground
amplitudes may slightly fluctuate over the mask. At this point, there originates from the
sequences. A convenient way from the point ISAR image synthesis module. The image
of view of the system operator; however, as formation process considers the images to be
shown the log-intensity domains of the two spatially periodic both in the horizontal and
classes are usually significantly overlapping, vertical, then, the imaging step estimates the
thus involving prior knowledge in the target center, and attempts to crop the
process may be necessary. The assumption appropriate rectangle of interest (ROI) from
of a prior estimation about the ratio of this periodic image .However, if the center
foreground areas compared with the image of the ROI is erroneously identified, the
size, which was a reasonable assumption target line segment may break into two (or
regarding large vessels, because the targets four) pieces. Therefore, in the proposed
showed line segment structure and the image processing approach, the longest
imaging step intended to provide us spatially foreground segment of the axis line in a
normalized images where the target is duplicated mosaic image, which steps also,
centered and the image is cropped so that it re-estimates the center of the input frame.
estimates a narrow bounding box of the
target. Thereafter, a preliminary foreground Scatterer Set Extraction
mask was derived by thresholding the input

All Rights Reserved © 2017 IJARMATE 70

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
Scattering mechanisms within the in the object configuration space. In the
same resolution cell and to defocusing following part of this section, two kernels
effects, the amplitudes may significantly are introduced and demonstrated their
vary over the consecutive frames, moreover effects. Thereafter, the details of the
we must expect notable differences between complete spatiotemporal model and the
different scatterers of the same frame. iterative optimization process are given. The
Permanent scatterers cause dominantly high input of the scatterer filtering kernel is the
amplitude. Demonstration of the actual estimation of the axis line segment
foreground–background segmentation has and the PSC set. The kernel exploits two
been represented. with background and facts observed in cases of large carrier ships:
foreground probability maps (high
probabilities indicated with greater 1) For a given target candidate, the
intensities). The foreground mask through scatterer candidates are choosen close
pixel-by-pixel maximum likelihood (ML) to the axis line.
classification based MRF optimization. The 2) The projection of two different
foreground mask by the proposed MRF scatterers to the axis line should not
model which effect is clearly demonstrated. be too close to each other, as the later
Thus, determination of efficient global artifacts are mainly caused by
thresholds to extract all scatterers by simple multiple echoes from the same
magnitude comparison. Therefore focusing scatterer. The minimal distance of
first on a high recall rate, a large group of two real scatterer projections in the
scatterer candidates is extracted, which may domain is determined by a threshold
contain several false positives. Thereafter, parameter, which varied between 0.05
an iterative solution to discriminate the real and 0.07.
scatterers from the false candidates, with
Based on the above assumptions, a
utilizing the temporal persistence of the
filtered scatterer set by a sequential
scatterer positions and the line structure of
algorithm is selected. However, this filter is
the imaged targets. A foreground constraint:
by nature very sensitive to the accuracy of
search for scatterers in the ISAR image
the preceding axis estimation step. As
regions labeled as by the initial input
shown here applying the SF kernel directly
binarization step. As the results in the real
for the output of the Hough based axis
scatterers are efficiently detected in this
detector results in notably weak
way, but the false alarm rate is high.
classification. For the above reason, a
Scatterer Filtering second move, which can be applied for
scenarios where a strong line-arrangement
The scatterer selection algorithm constraint is valid for the real scatterer
iterates various local moves, called kernels, configuration. The RANSAC algorithm after

All Rights Reserved © 2017 IJARMATE 71

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
obtaining a re-estimated axis, that the As the observed object’s structure can
RANSAC kernel has a couple of limitations: be considered rigid, a strong correlation
it cannot be adopted, if there are only a few between the target parameters in the
permanent scatterers on the target’s image consecutive frames received. Because of the
and the RANSAC-estimation may fail in imaging technique, the center is not relevant
cases of multiple duplicated scatterers in the regarding the real target position, thus only
PSC set that can form parallel lines. The penalize high differences between the angle
later artifact can appear as a consequence of and length parameters and significant
echoes in the imaging step. However, differences in the relative scatterer positions
assuming that the target structure is fixed, and scatterer numbers between close-in-time
and the position and orientation images of the sequence. The prior
displacement is small between consecutive interaction term is constructed as the
frames, temporal constraints can be weighted sum of four sub terms the median
exploited to refine the detector. angle difference ,the median scatterer
number difference and the median scatterer
Edge Preservation Of target Images alignment difference .

The Canny edge detection [22], which V. OPTIMIZATION

is commonly used to identify edge position. An image needs to be divided into clusters
It uses a smoothing process prior to edge without edges in order to prevent ringing
detection. The smoothing method often uses artifacts. However, the boundary definition
a Gaussian filter, but this distorts the edge of each region with uniform texture is
position if it is used in Canny edge difficult to provide using existing
detection. An edge-preserving smoothing segmentation methods. Thus, it is difficult to
method is thus needed. Patches that lie near determine the boundary of a region within a
the boundaries of a cluster or those that may noisy image. Optimizing is usually
have been misclassified. Hence, instead of performed with iterative stochastic
combining the different estimates of the algorithms, such as the reversible jump
same pixel, it simply retains the intensity Markov chain Monte Carlo sampler or the
estimate for which the pixel is at the center multiple birth and death dynamics.
of the patch being restored. The accuracy of However, in contrast to the above MPP
Canny edge detection by using bilateral models, in the proposed system, a single
filtering is determined [23],[24]. A bilateral object is tracked across several frames in the
filtering is an edge-preserving filter that uses input sequence, where the geometrical
weights defined according to the difference constraints are strong between the expected
of pixel luminance between the pixel of object parameters in near ISAR images. For
interest and its neighboring pixels. example, the length and orientation of the

All Rights Reserved © 2017 IJARMATE 72

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
imaged target cannot change significantly VI CONCLUSION
within a short observation period. As a In this method detection and tracking of the
consequence, the weight of the prior aircrafts and ships which are the moving
geometric terms increase and the iterative target in ISAR image sequences using
optimization process becomes sensitive to energy optimization process. By this the
be stuck in local. Solution is initialized with axis, the points and the moving target are
the output of the preliminary detector of extracted detected and tracked respectively.
which provides an initial configuration In the presence of noise sequences a multi
which is in most of the frames reasonably frame marked point process model is
consistent with the input data. Thereafter, introduced to remove the noise. Hence the
we proceed to an iterative refinement resulting target sequence is obtained by
algorithm, which scans in each step the iterative optimization process by
whole sequence and attempts to replace the simultaneously considering the observed
actual objects with more efficient ones image data and various geometric
considering the data and prior constraints in constraints. outperformed the other methods
parallel. The two key points of this with regard to sidelobe comparison but was
procedure are inferior to them with respect to modular
1) the generation step of new object index values. In addition, a comparison of
candidates the image showed that pixel differences
2) the evaluation step of the proposed yield a similar result to the threshold value.
objects
with regard to the current configuration and This indicates that denoising can be
the input data. For object generation, we use facilitated by image segmentation and edge
two types of moves: the perturbation Kernel preservation. The main limitation of this
and the RANSAC-based birth kernel, which work is that the performance of the proposed
are chosen randomly at each step of each method is dependent on the parameters in
iteration. The perturbation kernel clones the the partial processes. In future research, it is
actual object either from the current or the planned to replace edge-preserving
previous or the next frame; and it adds zero segmentation with a superior method. For
mean Gaussian random values to the center. instance, the edge regions and homogeneous
Finally, the scatterer positions are cloned regions should be defined using frequency
from the object of the current frame and domain parameters. Further, a method to
optionally additional scatterers are added or determine the threshold values is also to be
some scatterers are removed. considered.

All Rights Reserved © 2017 IJARMATE 73

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
REFERENCES Joint Conf. Neural Netw., vol. 1. 2000,
[1] 2246 ieee transactions on geoscience and pp. 163–168.
remote sensing, vol. 52, no. 4, april [8] Chatterjee.P and Milanfar.P,
2014. “Clustering-based denoising with locally
[2] Beck.A and Teboulle.M, “Fast gradient- learned dictionaries,” IEEE Trans. Image
based algorithms for constrained total Process., vol. 18, no. 7, pp. 1438–1451,
variation image denoising and deblurring Jul. 2009.
problems,” IEEE Trans. Image Process., Christoudias.C.M, Georgescu.B, and
vol. 18, no. 11, pp. 2419–2434, Nov. Meer.P, “Synergism in low level vision,”
2009. in Proc. 16th Int. Conf. Pattern
[3] Berizzi.F, Mese.E, Diani.M, and Recognit., vol. 4. 2002, pp. 150–155.
Martorella.M, “High-resolution ISAR [9] Christo Ananth, A.S.Senthilkani,
imaging of maneuvering targets by S.Kamala Gomathy, J.Arockia Renilda,
means of the range instantaneous ISAR G.Blesslin Jebitha, Sankari
ship images in 2011. @Saranya.S., “Color Image
[4] Buades.A, Coll.B, and Morel.J.M, “A Segmentation using IMOWT with 2D
review of image denoising algorithms, Histogram Grouping”, International
with a new one,” Multiscale Model. Journal of Computer Science and Mobile
Simul., vol. 4, no. 2, pp. 490–530, 2005. Computing (IJCSMC), Vol. 3, Issue. 5,
May 2014, pp-1 – 7
[5] Christo Ananth, G.Gayathri, I.Uma
Sankari, A.Vidhya, P.Karthiga, [10] Dabov.K, Foi.A, Katkovnik.V, and
“Automatic Image Segmentation method Egiazarian.K, “Image denoising by
based on ALG”, International Journal of sparse
Innovative Research in Computer and [11] 3-D transform-domain collaborative
Communication Engineering (IJIRCCE), filtering,” IEEE Trans. Image Process.,
Vol. 2, Issue 4, April 2014,pp- 3716- vol. 16, no. 8, pp. 2080–2095, Aug.
3721 2007.
[6] Cao et al.X., “Non-harmonic analysis [12] Descombes .X and Zerubia J, “Marked
applied to optical coherence tomography point processes in image analysis,” IEEE
imaging,” Jpn. J. Appl. Phys., vol. 51, Signal Process. Mag., vol. 19, no. 5, pp.
no. 2R, p. 022503, 2012. 77–84, Sep. 2002.
[7] Chatelain.F, Descombes.X, and [13] Dosselmann.R and Yang.X.D, “A
Zerubia.J, “Parameter estimation for in comprehensive assessment of the
Proc. IEEE Int. Radar Conf., Sep. 2003, structural similarity index,” Signal
pp. 36–41. ISAR images,” in Proc. Int. Image Video Process., vol. 5, no. 1, pp.
81–91, 2011

All Rights Reserved © 2017 IJARMATE 74

 ISSN (ONLINE): 2454-9762
ISSN (PRINT): 2454-9762
Available online at www.ijarmate.com

International Journal of Advanced Research in Management, Architecture, Technology and Engineering

(IJARMATE) Vol. 3, Special Issue 1, February 2017
[14] Elad.M and Aharon.M, “Image basis of spectral structure from 2-D
denoising via sparse and redundant nonharmonic analysis,” IEEE Trans.
representations over learned Image Process., vol. 22, no. 8, pp. 3008–
dictionaries,” IEEE Trans. Image 3017, Aug. 2013.
Process., vol. 15, no. 12, pp. 3736–3745, [18] Hirakawa.K and Parks.T.W, “Image
Dec. 2006. denoising using total least squares,”
[15] Foi.A, Katkovnik.V, and Egiazarian.K, IEEE Trans. Image Process., vol. 15, no.
“Pointwise shape-adaptive DCT for 9, pp. 2730–2742, Sep. 2006.
high-quality denoising and deblocking of [19] Ichinose.T, Hirobayashi.S, Misawa.T,
grayscale and color images,” IEEE and Yoshizawa.T, “Forecast of stock
Trans. Image Process., vol. 16, no. 5, pp. market based on nonharmonic analysis
1395–1411, May 2007. used on NASDAQ since 1985,” Appl.
[16] Guleryuz.O.G, “Weighted averaging for Financial Econ., vol. 22, no. 3, pp. 197–
denoising with overcomplete 208, 2012.
dictionaries,” IEEE Trans. Image [20] Katkovnik.V, Foi.A, Egiazarian.K, and
Process., vol. 16, no. 12, pp. 3020–3034, Astola.J, “From local kernel to nonlocal
Dec. 2007. multiple-model image denoising,” Int. J.
[17] Hasegawa.M,Kako.T, Hirobayashi.S.H, Comput. Vis., vol. 86, no. 1, pp. 1–32,
Misawa.T, Yoshizawa.T, and 2010.
Inazumi.Y, “Image inpainting on the

All Rights Reserved © 2017 IJARMATE 75