Segmentation of Complex Images using SOM for Clustering

1
Segmentation of Complex Images using SOM for
Clustering
Final Project for CS 7650 Instructor Dr. Heng-Da
Cheng

• Manasi Datar
• mpdatar_at_cc.usu.edu
• April 22, 2004

2
Presentation Outline

• Core Concepts
• Segmentation
• Segmentation and clustering
• Self-Organizing Map (SOM)
• Clustering and SOM
• Segmentation and SOM
• My Idea )
• Segmentation of complex images using SOM for
  clustering
• Approach not novel but applied to a new domain
  )
• Application pre-processing of images for
  knowledge discovery

3
Segmentation

• What is it?
• A process to distinguish meaningful objects in
  an image from the background
• How is it done?
• threshold techniques
• edge-based methods
• region-based techniques
• connectivity-preserving relaxation-based methods

4
Segmentation Techniques

• Threshold techniques
• Make decisions based on local pixel information
• Are effective when the intensity levels of the
  objects fall squarely outside the range of levels
  in the background
• Spatial information is ignored
• Edge-based techniques
• Center around contour detection
• Weakness connecting together broken contour
  lines
• Prone to failure in the presence of blurring

5
Segmentation Techniques contd

• Region-based Techniques
• Connected regions grouping neighboring pixels of
  similar intensity levels
• Region merging adjacent regions are merged under
  some criterion
• Overstringent criteria fragmentation
• Lenient blurred boundaries overlooked
  (overmerge)
• Connectivity-preserving Relaxation-based
  Techniques
• Active contour model
• start with some initial boundary shape and
  iteratively modify according to some energy
  function
• Getting caught in a local minimum is a risk

6
Segmentation and Clustering

• Clustering
• data set as a group of clusters collections of
  data points that belong together
• Segmentation as clustering
• represent an image in terms of clusters of pixels
  that belong together

7
Clustering Algorithms

• Make each point a separate cluster
• Until the clustering is satisfactory
• Merge the two clusters with the smallest
  inter- cluster distance
• End
• Agglomerative

• Construct a single cluster containing all points
• Until the clustering is satisfactory
• Split the cluster that yields the two
  components with the largest inter-cluster
  distance
• End
• Divisive

8
Clustering for Segmentation

• K-means algorithm
• A natural objective function can be obtained by
  assuming that we know there are k clusters, where
  k is known.
• Each cluster is assumed to have a center we
  write the center of the ith cluster as ci. The
  ith element to be clustered is described by a
  feature vector xi
• We now assume that elements are close to the
  center of their cluster, yielding the objective
  function
• Note if the allocation of points to clusters is
  known, it is easy to compute the best center for
  each cluster

9
Clustering for Segmentation contd

• K-means Algorithm
• Choose k data points to act as cluster centers
• Until the cluster centers are unchanged
• Allocate each data point to cluster whose center
  is nearest
• Now ensure that every cluster has at least one
  data point possible techniques for doing this
  include supplying empty clusters with a point
  chosen at random from points far from their
  cluster center
• Replace the cluster centers with the mean of the
  elements in their clusters
• End

10
Clustering for Segmentation

• k-means algorithm example
• On the left, an image of mixed vegetables, which
  is segmented using k-means to produce the images
  at center and on the right.
• Each pixel is replaced with the mean value of its
  cluster.
• In the center, a segmentation obtained using only
  the intensity information.
• At the right, a segmentation obtained using color
  information.
• Each segmentation assumes 5 clusters.

11
Structure of the SOM

• The SOM arranges feature vectors according to
  their internal similarity, creating a continuous
  topological map of the input space.
• In this topological map, the vectors that are
  similar in the input space are grouped together,
  or clustered.
• The map is a two dimensional grid of processing
  elements, called neurons.
• Each processing element is connected to each
  element of the input layer and to its neighboring
  processing elements.
• The connections between the layers (weights)
  represent the strength of the connection.

12
Working of the SOM

• The training algorithm is based on competitive
  learning
• only winning processing elements are allowed to
  adjust weights or learn
• After presenting an input vector, the processing
  elements response, or activation, is calculated
  by multiplying the input vector by a set of
  weights
• The processing element with the highest
  activation is declared winner
• The winning processing element and its neighbors
  are allowed to adjust their weights, while the
  remaining ones are left unchanged
• The neighborhood is the group of processing
  elements adjacent to the winning neuron
• During training the size of the neighborhood is
  reduced to stabilize the effect of the input
  vectors in the topological map
• The result of the training is a centroid
  (representative vector) for each cluster that
  minimizes the quantization error
• The quantization error is defined as the square
  root of the sum of the squared difference of
  individual input patterns and calculated centroids

13
SOM Demo

• I meant to show this in Matlab, but
  unfortunately this laptop does not have it
  installed(
• So, here is a Java version instead demo

14
SOM for Clustering
Input image
Topological map
15
SOM for clustering

• Properties of SOM
• Classification
• Visualization
• Can be used as a clustering tool )
• Will help in gaining useful information form
  segmenting complex images.

16
? My Idea ?

• Segmentation of Landsat-7 TM images using SOM
• Landsat information
• images freely available online
• Main reference
• Automatic Segmentation of Magnetic Resonance
  Images (MRIs) of the Human Brain using
  Self-Organizing Maps (SOMs) by Evangelou

17
Proposed Mechanism
Landsat 7 image (raw data)
Rectified Landsat 7 image
SOM
Corrections
Conventional Classifier
Classified image
Classified image
Post processing
Comparison and Assessment
18
Some Results
Original Image
Segmented Image
19
In Conclusion

• Results
• Comparison based on
• Time complexity
• Effectiveness
• SOM will be helpful in identifying intensity
  patterns that exist in image data
• Knowledge of such patterns will prove useful
  interpretation of such images

20
References

• Evangelou, I.E. Automatic Segmentation of
  Magnetic Resonance Images (MRIs) of the Human
  Brain using Self-Organizing Maps (SOMs). MS
  Thesis, Clarkson University, Potsdam, NY, USA
  (1999).
• Fayyad, U.M. Piatetsky-Shapiro, G. Smyth, P.
  and R. Uthurusamy, R., (Eds.). Advances in
  Knowledge Discovery and Data Mining. AAAI Press,
  the MIT Press, CA, USA (1996).
• Rangsanseri, Y et al. Multispectral Image
  Segmentation using ART1/ART2 Neural Networks.
  22nd Asian Conference on Remote Sensing,
  Singapore (2001)
• Solaiman, B Mouchot, M.C. A Comparative Study
  of Conventional and Neural Network Classification
  of Multispectral Data. IGARSS94, Pasadena, USA.
  (1994)
• Vesanto, J. Data Mining Techniques Based on the
  Self-Organizing Map. MSc Thesis, Helsinki
  University of Technology, Espoo, Finland (1997)
• Goldberg, M. Shlien, S. A clustering scheme for
  multi-spectral images. IEEE Transactions on
  Systems, Man and Cybernetics SMC 8. (1978)
• Theiler, J. Gisler, G. A contiguity-enhanced
  k-means clustering algorithm for unsupervised
  Multispectral image segmentation. Proc SPIE 3159.
  (1997)
• Landsat project info http//geo.arc.nasa.gov/sge/
  landsat/landsat.html

21
References

• Jain, A. K. Dubes, R. C. Algorithms for
  Clustering Data. Prentice-Hall Inc., Englewood
  Cliffs, NJ, USA. (1988)
• Haykin, S. Neural Networks A Comprehensive
  Foundation, Pearson Education (Singapore) Pte.
  Ltd. (2001)
• Brown, D. A. Craw I. Lewthwaite, J. Towards
  Core Image Processing with Self-Organizing Maps.
  (2001)
• Su, M. Clustering Analysis, Department of
  Computer Science and Information Engineering,
  National Central University (2003) available at
  http//selab.csie.ncu.edu.tw/muchun/course/cluste
  r/Clustering-crisp.pdf
• Mather P. Computer Processing of Remotely Sensed
  Images, John Wiley Sons, Inc. New York, NY, USA
  (1999)
• Paola, J. D. Schowenderdt, R. A. A review and
  analysis of back propagation neural networks for
  classification of remotely sensed multi-spectral
  imagery, International Journal of Remote Sensing.
  (1995)
• Landsat data http//landsat.gsfc.nasa.gov/data/Br
  owse/Cities/L7_CitiesGallery.html
• SOM toolbox for Matlab http//www.cis.hut.fi/proj
  ects/somtoolbox/

22
!!! Thank you for your patience !!! Your
comments and suggestions are valuable please
send them to me )