Sie sind auf Seite 1von 6

International Journal of Computer Science and Management Research Vol 1 Issue 4 November 2012

ISSN 2278-733X

Image Segmentation: A Review


Shraddha Tripathi Krishna Kumar B.K.Singh R.P.Singh
Electronics & Communication Engineering Department, BTKIT Dwarahat, Uttarakhand, India

Abstract- This paper presented a review on different image Thresholding is a very common approach used for Region-
segmentation techniques. Segmentation is the stage where a based segmentation where an image represented as groups of
significant commitment is made during automated analysis by pixels with values greater or equal to the threshold and values
delineating structures of interest and discriminating them from less to threshold value.
background tissue. It is also useful for feature extraction, image Clustering is also an approach for region segmentation where
measurements and image display. The aim of this paper is to give an image is partitioned into the sets or clusters of pixels
a detailed review on digital image segmentation techniques in having similarity in feature space. Region growing is another
three major fields of images as gray scale images, hyperspectral approach of region segmentation algorithms where assigned
images and medical images. It has been useful in medical image the adjacent pixels or regions to the same segment.
processing as well as hyperspectral image processing. This There are three types of images as gray scale, hyperspectral
evaluation on image segmentation techniques will be very useful and medical images.
for the development and improvement of new techniques.
II. SEGMENTATION FOR GREY SCALE IMAGES
Keywords- Segmentation, Medical Image Segmentation
Many techniques have been proposed to deal with the image
I. INTRODUCTION segmentation problem in Grey-Scale.
The image segmentation is a key process of the image analysis
and the image comprehension. Because of the influence of the (a) INVERSION TECHNIQUE
complicated background, the object characteristics diversity This technique is used to continuously update the muscle
and the noise, the image segmentation is the difficult and hot activity where produced a face movement following a given
research issues on the image processing. The process of face trajectory (Smith et al., 1986; Terzopoulos et al., 1991;
partitioning a digital image into multiple regions (sets of Press et al., 1992; Waters et al., 1996)[2]. When the inversion
pixels) is called image segmentation. Actually, partitions are had been carried out for all frames then the inverted activity
different objects in image which have the same texture or was generated an animation. A conventional nonlinear
colour. The result of image segmentation is a set of regions optimizer was used for minimizing a cost function that was
that collectively cover the entire image, or a set of contours selected to implement the inversion. The cost function E was
extracted from the image. All of the pixels in a region are the sum of the squares of the Euclidean distances between the
similar with respect to some characteristic or computed markers and the corresponding face model nodes:
property, such as color, intensity, or texture. Adjacent regions

are significantly different with respect to the same


characteristics. Some of practical applications of image E  |m  n |
segmentation are: image processing, computer vision, face 

recognition, medical imaging, digital libraries, image and


video retrieval, etc [1]. There are various segmentation Where m and n are the 3-D positions of the i marker and
techniques depending upon the classification schemes [1] face model node, respectively, N is the number of nodes used
i) Manual , semi-automatic and automatic in the inversion, and |m  n | is the vectorial magnitude
ii) Pixel-based (local methods), and Region-based square operator, i.e. the sum of the squares of each coordinate
(global methods) of the vector, so, that the inversion could produce different
iii) Manual delineation, low-level segmentation activity patterns, depend on the initial conditions. To limit the
(thresholding, region-growing, etc) and model-based number of solutions, some constraints added to the inversion.
segmentation (multispectral or feature map
techniques, dynamic programming, contour (b) ACTIVE CONTOUR MODELS
following, etc.) The goal of Active contour models, to apply segmentation
iv) Classical (thresholding, edge-based and region-based process to an image by deforming the initial contour towards
techniques), statistical, fuzzy and neural networks the boundary of the object. We do that by deforming an initial
techniques. contour to minimize the energy function which defined on
Basically the segmentation is divided in two categories- contours (Kass et al. 1987; Ray et al., 2002; Zimmer et al.,
1) Region-based segmentation and 2002; Kruse et al., 1996; Sacan et al., 2008) [2]. There are
2) Edge-based segmentation two components in this energy: the potential energy that is
small when the contour is aligned to the edges of the image,

Shraddha Tripathi et.al. 838 www.ijcsmr.org


International Journal of Computer Science and Management Research Vol 1 Issue 4 November 2012
ISSN 2278-733X

and the internal deformation energy, that is small for the


smooth contour. Both energy components are contour (f) GRAPH CUTS AND EFFICIENT N-D IMAGE SEGMENTATION
integrals with respect to a parameter of the contour. Active
contour models represented by two models: region based This approach is used to object/background segmentation
models and edge-based models. The characteristics of the using a s-t graph cuts. It is used to solve the problems in
image are used to determine the model. The advantage of this vision and graphics. The main goal is a completely automatic
model is to give a linear description of the object shape during high-level grouping of image pixels [5]. To build an efficient
the time of convergence without adding extra processing. object extraction tools for N-D applications where used a
binary graph-cuts based on contextual information and useful
(c)THRESHOLD METHOD topological constraint. The object segmentation is possible
Thresholding algorithms can be selected manually according only for some 2D object segmentation methods and are more
to a priori knowledge or automatically by image information. reliable and robust. The graph cuts formulated segmentation
These algorithms further divided to edge-based, region-based energy over binary variables whose values represent the
and hybrid. Edge-based algorithms are related with the edge location of the pixel whether inside or outside.
information. The Structures of an object can be depicted by
edge points. Common edge detection algorithms such as (g) K-WAY IMAGE SEGMENTATION TECHNIQUE FOR RANDOM
canny edge detector and Laplacian edge detector can be WALKS
classified to this type of regions. These algorithms are used to
find the edge pixels while eliminating the noise influence. For The K-way image segmentation technique which gives the
example, canny edge detector used the threshold of gradient user defined seeds where each seed specify the location
magnitude to find the potential edge pixels and suppressed belongs to K objects. Calculate the probability that walker
them through the procedures of the non maximal suppression first reaches each of the K-points [6]. Assign K-tuple vector to
and hysterics thresholding. As the operations used in these each pixel i.e. the probability of a random walker starting
algorithms are based on pixels, the detected edges are from each unseeded pixel will first reach each of the K-seed
consisted of discrete pixels and hence may be incomplete or points. The final segmentation derived from these K-tuples by
discontinuous. Hence, it must be apply the post processing selecting the most probable seed destination for a random
like morphological operation to connect the breaks or walker. Each edge assigned a real valued that corresponds to
eliminate the holes. This method has the ability that can be the probable that a random walker cross the edge. The
used to segment 3D image with good accuracy, but the advantage of formulating the probability on a graph made the
disadvantage of this method is the difficulty to process the operators used easily that require no discretization and
images of textured blob objects. therefore no error and ambiguities.

(d) EDGE FLOW METHOD (h) HYBRID IMAGE SEGMENTATION USING WATERSHED AND
It is used for boundary detection and required very little FAST REGION MERGING
parameter tuning. As the edges are located at local maxima There is a general segmentation problem as how to segment
by the gradient operator in the intensity feature. It first an image into homogeneous segments such that after
identified a flow direction at each pixel location where two combining two neighbors it gives a heterogeneous segment.
opposite direction of edge flow occurred. It is used to utilize There are many techniques for an error-free image partitions
the predictive coding model to identify and integrate the as histogram-based represents the simple probability
direction of change in some attributes such as colour, texture distribution function of intensity values of any image. Edge-
and phase discontinuities. based technique used to detect using differential filter in order
of image gradient or laplacian and then grouped them into
(e) CLUSTERING TECHNIQUE contours represents the surface. In the region-based
There are two approaches used in Clustering based segmentation technique segment the image into a set of
techniques- homogeneous regions then merged them according to certain
decision rules [7]. In the Markov random field based
i) K-means clustering-It is used to solve the low level
segmentation technique the true image is realized by a
segmentation tasks. This algorithm is convergent and Markov or Gibbs random field with a distribution function.
computationally efficient. Hybrid segmentation techniques are combined such as edge
ii) Fuzzy Clustering-It is effectively used in image processing based and region based techniques. In this image is firstly
pattern recognition and fuzzy modeling. There are different partitioned into regions and then merged them using split and
types of similarity measures used to identify classes depended merge technique and after that detected the contours using
on the data and the application. These similarity measures edge-based technique.
such as distance, connectivity, and intensity are used. Its
III. MEDICAL IMAGE SEGMENTATION
application is in data analysis, pattern recognition and image
Segmentation is an important tool in medical images analysis
segments. as it provides information about 3-D visualization, surgery. It

Shraddha Tripathi et.al. 839 www.ijcsmr.org


International Journal of Computer Science and Management Research Vol 1 Issue 4 November 2012
ISSN 2278-733X

is very useful in many applications including detection of the image and can be measured over spatial scales. B-mode
coronary border in angiograms, multiple sclerosis lesion ultrasound segmentation was used for similarity measure and
quantification, surgery simulations, surgical planning, provides a displacement estimate as constraint. The mutual
measuring tumor volume and its response to therapy, information of image geometry used for the evaluation of an
functional mapping, automated classification of blood cells, average intensity profile. The textures of an ultrasound image
studying brain development, detection of microcalcifications were dependent on the imaging system and characterized by
on mammograms, image registration, atlas-matching, the microstructure of the tissue being imaged.
detection of tumors, etc. Medical image segmentation iii) Shape- For a reliable and accurate segmentation the
algorithms have two components i.e. supervised and ultrasound images required not only the edge cues and region
unsupervised. information but also some shape constraints. The shape
Unsupervised algorithms are completely automatic and region constraints include the boundary regularization such that the
splitting with high density. It consists of different algorithms parametric shape and preferred shape was imposed.
as feature-space based technique, clustering based (K-means,
C-means, E-means), histogram thresholding, Image-domain or (c) WAVELET BASED SEGMENTATION OF HYPERSPECTRAL
Region-based techniques (Split-merge techniques, Region COLON TISSUE IMAGERY
growing, Neural-network techniques, Edge detection), Fuzzy
techniques, Wavelet based techniques etc. The hyperspectral imaging sensors are used to capture the
scenes in narrow spectral bands over visible and near infrared
(a) K-MEANS CLUSTERING AND MARKOV CONTROLLED wavelength range of electromagnetic spectrum. Their range
WATERSHED ALGORITHM was tens to hundreds of spectral features of the captured
material. The images are visualized as a 3D cube or a stack of
The watershed is used in the segmentation of medical images multiple 2D images where the cube face is represented in
using K-means clustering and Markov controlled watershed spatial co-ordinates fx, y and depth as a wavelength dλ.
algorithm. The watershed transform segmented the gray
matter, white matter and cerebrospinal fluid from magnetic
resonance (MR) brain images. Image gradient was used to
initial point and region could get by region growing in
watershed algorithm. A catchment basin represented the area
draining into a river or reservoir where bright areas were high
and dark areas were low that looked like a plane.
There are two approaches in terms of catchment basins and
watershed lines i.e. Rainfall approach and Flooding approach.
In rainfall local minima was assigned an exclusive tag. The
drop moved to low amplitude neighbor as long as it reached a
tagged pixel and tag value. Figure: 1 Hyperspectral image data representation
In flooding approach the intangible pixel holes are
corresponds to each local minima. The water entered into The colon of a human body is the upper part of the large
holes and filled the each catchment basin. intestine tube and the rectum is the lower part. They are
This algorithm has the advantage that it is a fast, simple and characterized as cancer instances and their composite name
intuitive method and produced an entire division of the image was Colorectal or Bowel cancer. The uncontrolled growth of
in regions even when the contrast is poor. The limitation on tissue cells in colon or rectum caused the colorectal cancer.
this approach was over-segmentation and sensitive to noise. The human colon tissue cell consists of four parts as nuclei,
cytoplasm, lamina, propria and lumen.
(b) ULTRASOUND IMAGE SEGMENTATION
There are some task-specific constraints or priors for a good
ultrasound image.
i) Speckle- Speckle simply represents the noise and there are
many algorithms proposed to reduce them where we measure
tissue displacement and optimize image registration. It gives
the granular appearance.
ii) Image Features-The Rayleigh distribution was used in
active contour approach and in the level set segmentation
algorithm. The most popular algorithm is the Gradient based
level set algorithm. We found the boundary or edge function
of the relative orientation of the boundary. For analyzing the Figure: 2 Human colons in Digestive system
good quality clinical images, intensity gradient information
was used. Phases are responsible to obtain the structure in an

Shraddha Tripathi et.al. 840 www.ijcsmr.org


International Journal of Computer Science and Management Research Vol 1 Issue 4 November 2012
ISSN 2278-733X

There are two segmentation ways as- The images contain the information formulated and
i) Spatial analysis-It exploits the spatial relationship. encapsulated.
ii) Spectral analysis-
The Wavelet multiresolution technique used to evaluate the
spatial relationship between the objects in an image at
multiple scale and falls in the former category. They are the
orthogonal basis function supported by time (space) and
represent the signal. Wavelet theory consists of the tools as
pyramidal image processing, subband coding and quadrature
mirror filtering. Textures represented by the attributes of
spatial arrangement of the gray levels of the pixels.

Figure: 4 BPT Constructions

There are many techniques employed for buildings and road


extraction such as Snakes employed by Cohen (1991).The
semi-automatic extraction of linear objects and roads by
Gruen and Li (1997) from remote sensing images. For colour
Figure: 3 Sequence of operations for Wavelet based texture information and localizing the road structures of small width
analysis given by Zafiroprulos and Schenk(1998). The limitation in
Snakes is to automatically change their topology during
IV. HYPERSPECTRAL IMAGE SEGMENTATION propagation given by Osher and Paragios (2003). This
The segmentation techniques represented by the integration of problem solved by Caselles (1997) by introducing geodesic
information from several images called as Multispectral or active contours (GAC). It further expanded and formulated by
Multimodal. In multispectral images where each pixel is Osher and Sethian (1998). Level set technique for
characterized by a set of features and the segmentation can be photogrammetric and remote sensing applications. In GAC,
performed in multidimensional feature space using clustering the initial curve was given before the curve proceeded to
algorithms. The hyperspectral image plot is the formation of evolution. Its energy depend on the edges (gradients)
three-dimensional feature space. There are many segmentation inheriting. In 2005, Cao et.al. employed an energy functional
techniques proposed as k-nearest neighbors, k-means, fuzzy c- based on Mumford –Shah segmentation model which was
means, artificial networks algorithm, binary partition tree, used to detect the automatic man-made object in aerial
level set segmentation, Spatial-Spectral graph, baysian images. The strategy used was the Coarse-to- fine and a
approach etc. fractal error metric. It is used not only for extracting the
(a) USING BINARY PARTITION TREE buildings but also to extract urban and sub-urban regions.
The hyperspectral images have led to the growing number of
applications such as remote sensing, food safety, healthcare or (c) ADAPTIVE MARKOV RANDOM FIELD FOR JOINT MIXING
medical research. In the pixel-spectrum based representation AND SEGMENTATION
using Binary Partition Tree (BPT). It stored the hierarchical Hyperspectral images are represented by a very high
region based representation in a tree structure. The resolution remote sensing images that required a hundred of
representation of many applications of tree is such as spectral bands but there is a problem of spectral unmixing.
segmentation, classification, and indexing, filtering, The unmixing algorithm is represented by the combination of
compression or object recognition. BPT branches which several pure material spectra or end members with relative
consist of analyzing its own BPT branches based on the fractions as abundances [12]. The endmember extraction
construction of the affinity matrices. It represents the distance algorithm (EEA) used to recover the spectral signature that
based measure according to the canonical correlations. The includes the minimum volume simplex analysis (MVSA) and
minimal cut proposed as BPT pruning by using the recursive N-FINDR algorithm. The Baysian approach where spatial
spectral graph partitioning algorithm. Correlation between pixels was taken by using Markov-
There are three types of nodes such as leaf node, root node random Fields (MRF) to model pixel dependencies resulted in
and merging the neighbouring nodes. a joint segmentation and unmixing algorithm. The MRF
actively used for modeling spatial correlations in image
(b) A REGION BASED LEVEL SET SEGMENTATION [13][14]. The drawback of this technique is the computational
This segmentation is used for the road detection, buildings cost proportional to the size of the image. This problem
detection and other man-made objects in the satellite images. reduced by using the mathematical morphology based on

Shraddha Tripathi et.al. 841 www.ijcsmr.org


International Journal of Computer Science and Management Research Vol 1 Issue 4 November 2012
ISSN 2278-733X

lattice theory where similar regions are built by using filter ii) The segmentation is a very challenging task in
stemming. hyperspectral images irrespective of the given input. It is a
very complicated task to manage the amount of information
(d) ROAD EXTRACTION USING SVM that a hyperspectral data have. The dimensionality reduction
Road information is an important data layer in the technique is used to avoid the redundancies and irrelevant data
Geographical information System (GIS). The automatic road in terms of computational or time-saving or accuracy. The
extraction save time and labor. There are several Methods for texem model was applied on color images for segmentation.
road extractions as- The labeled data are hard to obtain due to the high variability
i) Ridge Finding- In this technique edge magnitude and of the input images so here used the unsupervised way.
direction are devised by edge operators followed by
thresholding and thinning process to get the ridge pixels given (f) CLUSTERING-BASED BAND SELECTION
by Nevatia and Babu in 1980 and Treash and Amaratunga in The drawback of the hyperspectral images are-
2000. These ridge pixels are linked to get the road segment. i) It has a large amount of information hence It is very
ii) Heuristic reasoning is based on knowledge of a series of a difficult to the application of supervised techniques.
pre-set rules on road characteristics as shape index, the ii) For the classification or representation task it is represented
distance between image primitives, fragments trend and in a large number of bands that are highly correlated.
contextual information to detect and connect the primitives or To reduce the hyperspectral representation we used feature
antiparallel linear edges to road segments by McKeown in extraction and feature selection techniques. Feature extraction
1985 and Zhu and Yen in 1986. In the dynamic programming represented by the transformed initial information to obtain
methods used a model of the roads with a set of mathematical the new and reduced data set. Feature selections have a subset
equation of roads like smooth curves, homogeneous surface, of relevant data from the original information. Hence the
narrow linear features and relatively constant width. The processing step in hyperspectral imaging on pixel
statistical interference method used to model the linear classification tasks performed by the redundant information
features as a Markov point process or a geometric-stochastic without loosing accuracy and no supervised information.
model on the road width, direction, intensity, background
intensity and maximum a posteriori probability used to (g) BAYSIAN HYPERSPECTRAL IMAGE SEGMENTATION WITH
estimate the road network given by Barzohar and Cooper in DISCRIMINATIVE CLASS LEARNING
1996 and Stoica et.al. in 2000. In map-matching method,
existing road maps were used as the starting point and update The applications of hyperspectral images are very dedicated
the road network. task as it is difficult to learn high dimensional densities by a
Support vector Machine (SVM) involves two stapes very limited number of training samples. The discriminative
i) A roads group that include actual roads and features having approach deals with small class distances, high dimensionality
some spectral reflectance. and limited training sets. It holds the state-of art in supervised
ii) A non- road groups that include all other features having hyperspectral image [17].
spectral values different from roads. The Baysian segmentation approach for hyperspectral images
The SVM is used for the bias- variance trade- off and over- where spatial dependencies enforced by a multi-level logstic
fitting problem by achieving the best generalization (MLL), Markov-Gibs prior in favor of neighboring labels of
performance where balancing the relationship between the the same class. These class densities built on the
accuracy obtained on the training data and the capacity of the discriminative Fast Sparse Multiresolution Regression
machine given by Vapnik at AT&T Bell Laboratories in 1995, (FSMIR) which has a fast version of the sparse multinomial
Then applied in every fields of image processing as regression (SMLR). The SMLR includes Laplacian prior to
handwritten digit recognition, face detection in images and control the complexity of the classifier and achieve good
text categorization. It is used to separate the classes so as to generalization capabilities.
maximize the margin in a hyperplanes surface also
implemented the structural risk minimization principle by (h) SPECTRAL–SPATIAL CLASSIFICATION OF HYPERSPECTRAL
Vapnik in 1982. DATA BASED ON A STOCHASTIC MINIMUM SPANNING FOREST
APPROACH
(e) UNSUPERVISED IMAGE SEGMENTATION BASED ON TEXEMS The recent algorithm was proposed as Minimum Spanning
The hyperspectral images having the several emerging Forest (MSF) approach which is used for the best
applications in the fields such as remote sensing, fine arts etc. classification accuracy values by automatically selected
Segmentation techniques are used to detect the data of salient markers [20]. The markers represented by the pixels that are
regions as accurate as possible. initiated the growing procedure where each marker being
There are two problems in the hyperspectral image assigned a distinct label. The seeds are chosen to initiate the
segmentation technique- growing procedure of the forest i.e. crucial. The pixel wise
i) The hyperspectral images involved a large amount of classification of the hyperspectral data set is performed and
information and had the problem of how to select the useful computed the probabilistic map. From the classification map,
information by avoiding redundancies. the most reliable pixels are kept as markers using which the

Shraddha Tripathi et.al. 842 www.ijcsmr.org


International Journal of Computer Science and Management Research Vol 1 Issue 4 November 2012
ISSN 2278-733X

MSF is built. This method was used to provide very good [18] Boykov, Y., Veksler, O., Zabih, R “Fast Approximate Energy
Minimization via Graph Cuts” IEEE Transactions on Pattern Analysis
resulting in terms of classification accuracy. The generation of
and Machine Intelligence, Vol. 23,Issue 11, pp.1222-1239, 2001
this map is very demanding in terms of algorithm complexity [19] Jun Li, José M. Bioucas-Dias and Antonio Plaza “Hyperspectral Image
and having low cost. Segmentation Using a New Bayesian Approach With Active Learning”
CONCLUSIONS IEEE Transactions On Geoscience And Remote Sensing, Vol. 49, No.
This paper presented a review of different segmentation 10, October 2011.
[20] Jun Li, José M. Bioucas-Dias, and Antonio Plaza “Spectral–Spatial
techniques used in the three major fields as Grey-scale Hyperspectral Image Segmentation Using Subspace Multinomial
images, medical images and hyperspectral images. It is very Logistic Regression and Markov Random Fields” IEEE Transactions on
important and useful to analyze all the techniques for future Geoscience and Remote Sensing, Vol. 50, No. 3, March 2012.
purposes and inventions of new techniques in each field. [21] Coskun, H., Li, Y., and Mackey, M. A.. Ameboid cell motility: A
model and inverse problem, with an application to live cell imaging
data. Journal of Theoretical Biology, 244(2):169–179,2007.
REFERENCES [22] S. Osher and J. Sethian, “Fronts propagating with curvature dependent
[1] Isaac N. Bankman “Handbook of medical imaging processing and speed: Algorithms based on Hamilton - jacobi formulations,” Journal of
analysis” Acadmic Press, 2000. Computationl Physics, pp. 12–49, 1988.
[2] Ashraf A. Aly, Safaai Bin Deris and Nazar Zaki “Research Review For [23] C. A. Glasbey, “An analysis of histogram-based thresholding
Digital Image Segmentation Techniques” International Journal of algorithms,”Comput. Vis. Graph. Image Process., vol. 55, pp. 532–
Computer Science & Information Technology (IJCSIT), Vol 3, No 5, 537,1993.
Oct 2011. [24] N. Otsu, “A threshold selection method from gray-level
[3] Wei-Ying Ma and B. S. Manjunath “Edge Flow: A Technique for histograms,”IEEE Trans. Syst., Man, Cybern., vol. SMC-9, no. 1, pp.
Boundary Detection and Image Segmentation” IEEE Transactions On 62–66, 1979.
Image Processing, Vol. 9, No. 8, August 2000 [25] R. Adams and L. Bischof, “Seeded region growing,” IEEE Trans.
[4] K.Muthukannan and P.Latha “Clustering Techniques based Crops Pattern Anal. Machine Intell., vol. 6, June 1994.
Image Segmentation” International Conference on Recent Trends in [26] S. Kumar, J. Ghosh, and M. M. Crawford, “Best-bases feature extraction
Computational Methods, Communication and Controls, 2012. algorithms for classification of hyperspectral data,” IEEE Trans. Geosci.
[5] Yuri Boykov and Gareth Funka-Lea “Graph Cuts and Efficient N-D Remote Sens., vol. 39, no. 7, pp. 1368–1379, Jul. 2001
Image Segmentation” International Journal of Computer Vision 70(2), [27] Agouris, P., S. Gyftakis, and A. Stefanidis.” Dynamic node distribution
pp.109–131, 2006 in adaptive snakes for road extraction,” Vision Interface, Ottawa,
[6] Leo Grady “ Random Walks for Image Segmentation” IEEE Canada, pp. 134–140,2001.
Transactions On Pattern Analysis And Machine Intelligence, Vol. 28, [28] D. Landgrebe, Signal Theory Methods in Multispectral Remote Sensing.
No. 11, November 2006. Hoboken, NJ: Wiley, 2003.
[7] Kostas Haris, Serafim N. Efstratiadis, Nicos Maglaveras, and Aggelos [29] G. Camps-Valls and L. Bruzzone, “Kernel-based methods for
K. Katsaggelos “Hybrid Image Segmentation Using Watersheds and hyperspectral image classification,” IEEE Trans. Geosci. Remote Sens.,
Fast Region Merging” IEEE Transactions On Image Processing, Vol. 7, vol. 43, no. 6, pp. 1351–1362, Jun. 2005.
No. 12, December 1998. [30] H. Bagan, Y. Yasuoka, T. Endo, X. Wang, and Z. Feng, “Classification
[8] K. Hohne, H. Fuchs, and S. Pizer, “3D Imaging in Medicine: of airborne hyperspectral data based on the average learning subspace
Algorithms, Systems, Applications” Berlin, Germany: Springer-Verlag, method,” IEEE Geosci. Remote Sens. Lett., vol. 5, no. 3, pp. 368–372,
1990. Jul. 2008.
[9] J. M. Gauch and S. M. Pizer “Multiresolution analysis of ridges and
valleys in gray-scale images” IEEE Trans. Pattern Anal. Machine Intell.,
vol. 15, pp. 635–646, June 1993.
[10] Chunming Li, Rui Huang, Zhaohua Ding, J. Chris Gatenby, Dimitris
N. Metaxas, and John C. Gore “A Level Set Method for Image
Segmentation in the Presence of Intensity Inhomogeneities With
Application to MRI” IEEE Transactions On Image Processing, Vol. 20,
No. 7, July 2011.
[11] J. Alison Noble and Djamal Boukerroui “Ultrasound Image
Segmentation: A Survey” IEEE Transactions On Medical Imaging, Vol.
25, No. 8, August 2006
[12] N. Keshava and J. Mustard, “Spectral unmixing,” IEEE Signal
Processing Magazine, pp. 44–56, Jan. 2002.
[13] F. Arduini, C. Dambra, and C. S. Regazzoni, “A coupled MRF model
for SAR image restoration and edge-extraction,” in Proc. IEEE Int.
Conf. Geoscience and Remote Sensing, vol. 2, pp. 1120–1122, 1992.
[14] K. Held, E. Kops, J. Krause, W. Wells, R. Kikinis, and H. Muller-
Gartner, “Markov random field segmentation of brain MR images,”
IEEE Trans. Med. Imag., vol. 16, no. 6, pp. 878–886, Dec. 1997.
[15] Julio M. Duarte-Carvajalino, Guillermo Sapiro,Miguel Vélez-Reyes,
and Paul E. Castillo “Multiscale Representation and Segmentation of
Hyperspectral Imagery Using Geometric Partial Differential Equations
and Algebraic Multigrid Methods” IEEE Transactions On Geoscience
And Remote Sensing, Vol. 46, No. 8, August 2008.
[16] Adolfo Martínez-Usó, Filiberto Pla, José Martínez Sotoca, and Pedro
García-Sevilla “Clustering-Based Hyperspectral Band Selection Using
Information Measures” IEEE Transactions On Geosciences and Remote
Sensing, Vol. 45, No. 12, December 2007.
[17] Janete S. Borges, José M. Bioucas-Dias and Andre R. S. Marcal
“ Bayesian Hyperspectral Image Segmentation with Discriminative
Class Learning.” IEEE Transactions on Geoscience And Remote
Sensing, Vol. 49, No. 6, June 2011.

Shraddha Tripathi et.al. 843 www.ijcsmr.org

Das könnte Ihnen auch gefallen