Docstoc

cbir

Document Sample
cbir Powered By Docstoc
					Extraction of major object features using VQ
clustering for content-based image retrieval

   Authors: Hun-Woo Yoo, She-Hwan Jung,
                 Dong-Sik Jang, Yoon-Kyoon Na
   Source: Pattern Recognition 35 (2002)
    1115-1126
   Speaker: Yu Yuan-Hui
   Date: 2002/6/20

     2002/6/20                                  1
                 Outline

   Introduction
   Image features extraction
   Representative feature extraction
   Image retrieval using object features
   Experimental results
   Conclusions



    2002/6/20                           2
              Introduction

       An image representation method
        using VQ on color and texture is
        proposed.

       The proposed method also
        provided for the object-based
        image retrieval.




2002/6/20                                  3
          Image features extraction
                     Acquiring raw image data (RGB information)


Transformation (RGB to HSV)                   Transformation (RGB to GRAY)


                                             Gray-level co-occurrence matrix
 Color texture extraction

          Hue,                                  Texture feature extraction
       Saturation,
         Value                                  Angular second moment,
                                                 Contrast, Correlation,
                                                   Variance, Entropy

                                               Assigning block-based textural features
                                                   to pixel-based textural feature


     Color and textural features in each pixel (8-dimensional vector)

   2002/6/20                                                                             4
        Representative feature extraction
        Input    Color and textural features in each pixel (8-dimensional vector)

                                                        VQ clustering

        Output         Representative feature of objects in the image)


                                      Object                    Sky      House    Straw    Grass
                                      Cluster                   1        2        3        4
                                      Cluster members           748      422      370      290
                                                Hue             0.5653   0.0576   0.0904   0.3319
                                                Saturation      0.3820   0.3025   0.4357   0.2440
                              Sky               Value           0.4300   0.4014   0.6291   0.3396
House                                 Centroid ASM              0.0151   0.0017   0.0013   0.0030

                              Straw             Contrast        0.0065   0.0294   0.0399   0.0193
Grass                                           Correlation     0.7690   0.6932   0.5969   0.6505
                                                Variance        0.0256   0.0995   0.1006   0.0601
                                                Entropy         0.4184   0.5901   0.6014   0.5779

           2002/6/20                                                                                5
Image retrieval using object features
          (0.422, 0.2366, 0.5722, 0.0019, 0.0054, 0.5864, 0.0156, 0.5762)



                                       0.31
                                      0.30
                                       0.52
                                       0.87

                   Query Image               Database Image

dQD = w9[w1(HQ-HD)2+w2(SQ-SD)2+w3(VQ-VD)2]1/2

     + w10[w4(ASMQ-ASMD)2
                               W1     W2      W3     W4    W5    W6    W7    W8    W9    W10
     + w5(CONTQ-CONTD)2
                               0.44   0.28    0.28   0.2   0.2   0.2   0.2   0.2   0.5   0.5
    + w6(CORQ-CORD)2

     + w7(VARQ-VARD)2

     + w8(ETRPQ-ETRPD)2]1/2
     2002/6/20                                                                           6
            Experimental results




2002/6/20                          7
2002/6/20   8
            Table 2 Comparing clustering time                                        Table 3 Retrieval rate

     The number        VQ(s)            EM(s)                   The number of          The number of relevant images
     of images                                                  Retrieved images
     100               139.0            152.9                                          C:1.0      C:0.0       C:0.5          C:0.7
                                                                                       T:0.0      T:1.0       T:0.5          T:0.3
     200               245.5            305.4
                                                                5                      3.8        3.8         4.1            3.8
     300               336.7            462.6
                                                                10                     6.0        5.1         6.9            6.2
     400               467.9            614.9
                                                                15                     7.2        6.2         9.2            8.9
     500               556.5            766.3

     600               647.4            915.4                   20                     9.1        8.8         11.7           11.2



                                        Table 4 Retrieval results with six object queries

The number                         The number of                                                Retrieval rates
of retrieved
images          Bear     Person   Sunset        Rose    Horse        Valley   Bear   Person     Sunset    Rose        Horse         Valley


5               4.8      4.7      4.7           5.0     4.1          4.8      0.96   0.94       0.94      1.0         0.82          0.96

10              8.8      8.3      8.0           9.4     6.9          9.3      0.88   0.83       0.80      0.94        0.69          0.93

15              12.3     11.4     10.6          13.5    9.2          13.2     0.82   0.76       0.71      0.90        0.61          0.88

20              14.9     13.7     12.1          14.9    11.7         15.8     0.75   0.69       0.61      0.73        0.59          0.79


           2002/6/20                                                                                                                   9
                 Conclusions

   Experiments showed 0.59-0.79 retrieval
    rates for selected query objects in case of
    asking 20 retrieved images.
   To search for other useful features.




     2002/6/20                                    10

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:5
posted:2/9/2012
language:English
pages:10