Docstoc

2007-Image segmentation by histogram thresholding

Document Sample
2007-Image segmentation by histogram thresholding Powered By Docstoc
					Image segmentation by histogram thresholding
using hierarchical cluster analysis
      Source: Pattern Recognition Letters, VOL. 27,
               Issue 13, October 2006
      Authors: Agus Zainal Arifin, Akira Asano
      Speaker: Pei-Yen Pai
      Date: 2007.05.10
                    Outline
•   Introduction
•   Otsu’s method
•   Proposed method
•   Experiment results
•   Conclusions
                      Introduction           Th1      Th2

  • Image segmentation           Pixel
                                 Count
     – thresholding
                                         0                  255
                                             Gray-level




Original image        Thresholded image        Contour image
                      Otsu’s method
  • The most common used thresholding method.
  • Simplicity and efficiency.
  • Maximize between-class variance or Minimize
    within-class variance.
σBetween(t )  PC1 (t )(MC1 (t )  M )2  PC2 (t )(MC2 (t )  M )2
 2


 Pci: The probability of i-th class.
 Mci: The mean of i-th class.
 M: The mean of image.
         Drawback of Otsu’s method
                             Th1      Th2
                 Pixel
                 Count


                         0                  255
                             Gray-level




Original image       Thresholded image            Contour image
             The proposed method




Histogram of the sample image   The obtained dendrogram
                    The proposed method

Inter-class                         Intra-class




              Ck1       Ck2                       Ck3
      0                                                 255
                              Gray-level
              The proposed method



    Dist 1 Dist 2 Dist 3

2        3     4       5       150   200
         The proposed method
The pair of the smallest distance is Dist 2




      Dist 1’      Dist 2’


  Dist 1 Dist 2        Dist 3


  2        3       4         5                150   200

               Merge
             The proposed method
 Dist A < Dist B


         Three groups Two groups




Dist A                             Dist B




  2      3         50         75            150   200
Experiment results



       Original images




 The histogram of Original images
   Experiment results



The thesholded images by proposed method




The thesholded images by Otsu’s method
  Experiment results



The thesholded images by KI’s method




The thesholded images by Kwon’s method
   Experiment results



The thesholded images by proposed method




The ground-truth of original images
Experiment results
                 Conclusions
• Present a new gray level thresholding algorithm.
• The proposed thresholding method yields better
  images, than those obtained by the widely used
  Otsu’smethod, KI’s method, and Kwon’s

				
DOCUMENT INFO