43
1 A Briefly Introduction of Region-Based Image Segmentation Advice Researcher: 丁丁丁丁丁 (Jian-Jiun Ding ) Presenter: 丁丁丁 (Cheng-Jin Kuo) Digital Image and Signal Processing Laboratory Graduate Institute of Communication Engineering National Taiwan University, Taipei, Taiwan, ROC 台台台台台 台台台台台台台台台台台台

A Briefly Introduction of Region-Based Image Segmentation.ppt

  • Upload
    volien

  • View
    214

  • Download
    0

Embed Size (px)

Citation preview

1

A Briefly Introduction of Region-Based Image Segmentation

Advice Researcher: 丁建均教授 (Jian-Jiun Ding )Presenter: 郭政錦 (Cheng-Jin Kuo)

Digital Image and Signal Processing LaboratoryGraduate Institute of Communication EngineeringNational Taiwan University, Taipei, Taiwan, ROC台大電信所 數位影像與訊號處理實驗室

2

Outline Introduction Data Clustering Hierarchical Clustering Partitional Clustering (k-means method) Simulation Results Comparison Conclusion and Future Work References

3

Introduction

Image Segmentation:

Threshold Technique Edge-Based Segmentation

Watershed Algorithm Method Region-Based Segmentation

Data Clustering Region Growing Region Splitting and Merging

4

Introduction

Image Segmentation:

Threshold Technique Edge-Based Segmentation

Watershed Algorithm Method Region-Based Segmentation

Data Clustering Region Growing Region Splitting and Merging

5

Data Clustering

Data Clustering:

Goal: 1. To classify the statistical databases 2. To segment the digital images by a systematical way.

6

Data Clustering

Two Kinds of Data Clustering :

Hierarchical Clustering hierarchical agglomerative algorithm hierarchical divisive algorithm

Partitional Clustering K-means method

7

Data Clustering

Two Kinds of Data Clustering :

The biggest difference between them: Hierarchical Clustering

The number of clusters is flexible. Partitional Clustering

The number of clusters is assigned before processing.

8

Data Clustering

Two Kinds of Data Clustering :

Hierarchical Clustering hierarchical agglomerative algorithm hierarchical divisive algorithm

Partitional Clustering K-means method

9

Hierarchical Clustering

Characteristics:Has been widely used in statistical fieldsTree DiagramSimple ConceptUser could change the number of clusters during

the processing timeWhich means: the numbers of clustering is flexible

10

Hierarchical Clustering

hierarchical agglomerative algorithm :

1. Treat every single point (pixel) as a cluster.

Numbers of cluster: 9

11

Hierarchical Clustering

hierarchical agglomerative algorithm :

2. Find out two clusters ,

for that:iC jC

( , ) m in ( ( , )),

x yi j x yD C C D C C

for

12

Hierarchical Clustering

hierarchical agglomerative algorithm : The Definition of Distance D:

Single-Linkage agglomerative algorithm : Complete-Linkage agglomerative algorithm :

( , ) min( ( , )), a , bi j i jD C C d a b for C C

( , ) max( ( , )), a , bi j i jD C C d a b for C C

13

Hierarchical Clustering

hierarchical agglomerative algorithm :

3. Agglomerate them together to form a new cluster.

Numbers of cluster: 8

14

Hierarchical Clustering

hierarchical agglomerative algorithm :

4. Repeat step1~step3 until the number of cluster satisfies the requirement.

15

Hierarchical Clustering

hierarchical agglomerative algorithm :

4. Repeat step1~step3 until the number of cluster satisfies the requirement.

Numbers of cluster: 7

16

Hierarchical Clustering

hierarchical agglomerative algorithm :

4. Repeat step1~step3 until the number of cluster satisfies the requirement.

Numbers of cluster: 6

17

Hierarchical Clustering

hierarchical agglomerative algorithm :

4. Repeat step1~step3 until the number of cluster satisfies the requirement.

Numbers of cluster: 5

18

Hierarchical Clustering

hierarchical agglomerative algorithm :

4. Repeat step1~step3 until the number of cluster satisfies the requirement.

Numbers of cluster: 4

19

Hierarchical Clustering

hierarchical agglomerative algorithm :

4. Repeat step1~step3 until the number of cluster satisfies the requirement.

Numbers of cluster: 3

20

Hierarchical Clustering

hierarchical agglomerative algorithm :

4. Repeat step1~step3 until the number of cluster satisfies the requirement.

Numbers of cluster: 2

21

Hierarchical Clustering

The tree diagram: (an example)

These two tree diagrams are downloaded from http://neural.cs.nthu.edu.tw/jang/ Author: Professor Jyh-Shing Roger Jang

22

Hierarchical Clustering

Advantages:Simple: We can check all the stages by checking out the tree

diagram only.Number of cluster is flexible: We can change the number of cluster anytime

during processing. which means?

23

Hierarchical Clustering

Drawback:

Slower compared to partitional clustering: Not suitable for processing larger database

24

Hierarchical Clustering

Two Kinds of Data Clustering :

Hierarchical Clustering hierarchical agglomerative algorithm hierarchical divisive algorithm

Partitional Clustering K-means method

25

Hierarchical Clustering

hierarchical division algorithm :

1. Treat the whole database as a cluster.

Numbers of cluster: 1

26

Partitional Clustering

Characteristics:

Computational time is short User have to decide the number of clusters before

starting classifying data The concept of centroid One of the famous method: K-means Method

27

Partitional Clustering

K-mean algorithm :

1. Decide the number of the final classified result with N.

Numbers of cluster: Nwe now assume N=3

28

Partitional Clustering

K-mean algorithm :

2. Randomly choose N point for the centroids of cluster.

(N=3)Numbers of cluster: N

29

Partitional Clustering

K-mean algorithm :

3. Find the nearest point for every centroid of cluster. Classify the point into the cluster.

Notice the definition of the nearest!

30

Partitional Clustering

K-mean algorithm :

4. Calculate the new centroid of every cluster.

Notice the definition of the centroid!

31

Partitional Clustering

K-mean algorithm :

5. Repeat step1~step4 until all the point are classified.

32

Partitional Clustering

K-mean algorithm :

5. Repeat step1~step4 until all the point are classified.

33

Partitional Clustering

K-mean algorithm :

5. Repeat step1~step4 until all the point are classified.

34

Partitional Clustering

K-mean algorithm :

5. Repeat step1~step4 until all the point are classified.

35

Partitional Clustering

K-mean algorithm :

5. Repeat step1~step4 until all the point are classified.

36

Partitional Clustering

K-mean algorithm :

6. Data clustering completed

N=3

37

Partitional Clustering

Advantages:Fast: Since the N has been decided at beginning, the

computational time is short.Simple: Since the N has been decided at beginning, every

simulation testing will be clear.

38

Partitional Clustering

Drawbacks:N is fixed: What is the best value of N?Choice of initial points influence the resultCircular shape

39

Simulation Results

Original Image

N=2 N=3

N=4 N=5

Processed by K-means:

40

Comparison

Hierarchical clustering:ReliableBut is time consumingCannot perform on a large database

Partitional clustering:FastBut got the initial problemAnd the circular shape problem

41

Conclusion and Future Work

K-means algorithm has been widely used in image segmentation field because of the fast processing speed.

How do we choose the best N? Can we use PSO algorithm to improve it? Other ideas of image segmentation:

Segment the image by the graphical features

42

References R. C. Gonzalez, R. E. Woods, Digital Image Processing 2nd Edition, Prentice Hall,

New Jersey 2002. M. Petrou and P. Bosdogianni, Image Processing the Fundamentals, Wiley, UK, 2004. W. K. Pratt, Digital Image Processing 4nd Edition, John Wiley & Sons, Inc., Los Altos,

California, 2007. S.C. Satapathy, J. V. R. Murthy, R. Prasada, B.N.V.S.S, R. Prasad, P.V.G.D. “A

Comparative Analysis of Unsupervised K-Means, PSO and Self-Organizing PSO for Image Clustering”, Conference on Computational Intelligence and Multimedia Applications, 2007. International Conference on Volume: 2  13-15 Dec. 2007, Page(s): 229-237

J. Tilton, “Analysis of Hierarchically Related Image Segmentations”, pre-sented at the IEEE Workshop on Advances in Techniques for Analysis of Remotely Sensed Data, Greenbelt, MD, 2003.

43

THANK YOU