80 Sep CBIR 6),7) WWW WWW Image Collector Image Collector (1) (2) 1 WWW 2 CBIR WWW WWW WWW CBIR example-based generic object recognition 8),9) W

Size: px

Start display at page:

Download "80 Sep CBIR 6),7) WWW WWW Image Collector Image Collector (1) (2) 1 WWW 2 CBIR WWW WWW WWW CBIR example-based generic object recognition 8),9) W"

ひでかかいて
5 years ago
Views:

1 Vol. 42 No. SIG 10(TOD 11) Sep WWW WWW World-Wide Web WWW WWW WWW Image Collector Image Collector (1) (2) 1 WWW An Image-gathering System from WWW Employing Keywords and Image Features Keiji Yanai Due to the recent explosive progress of WWW (World-Wide Web), we can easily access a large number of images over WWW. There are, however, no established methods to make use of WWW as a large image database. In this paper, we propose an automatic image-gathering system from WWW employing both keywords and image features, which is called the Image Collector. In our system, since image files on WWW are linked by HTML documents, at first, using keyword-based WWW search engines, we access and analyze a lot of HTML documents related to query keywords given by a user, and we fetch only images strongly related to the keywords. We divide fetched images into group A, in which images can be regarded as almost correct images only by analysis of HTML documents, and group B, in which image features of images need to be examined. By selecting large clusters from a clustering result of group A and selecting images from group B that are similar to selected images from group A, we eliminate images unrelated to the keywords, and finally, we get images related to the keywords. In the experiments, we obtained more than one hundred images in about five minutes. 1. WWW World Wide Web WWW WWW WWW WWW Department of Computer Science, University of Electro-Communications WWW Lycos Multimedia Search 1) AltaVista Image Search 2) Google Image Search 3) Content-Based Image Retrieval CBIR 4),5) 79

2 80 Sep CBIR 6),7) WWW WWW Image Collector Image Collector (1) (2) 1 WWW 2 CBIR WWW WWW WWW CBIR example-based generic object recognition 8),9) WWW 10) 2. WWW WebSeer 11) WebSEEk 12) Image Rover 13) 3 3 WWW HTML HTML HTML HTML WebSeer 11) WebSeer Face Detector 14) WebSEEk 12) CBIR 2 VisualSEEk 15) Image Rover 13) WebSEEk 2 HTML WebSEEk WebSEEk Image Rover WebSeer Face Detector

3 Vol. 42 No. SIG 10(TOD 11) WWW 81 3 Bianchi 16) Inder 17) K-DIME 3. Image Collector Image Collector WWW 2 (1) WWW (2) WWW WWW 1 WWW (1) Web URL Universal Resource Locator (2) URL Web Web HTML (3) HTML HTML HTML A B C (4) A B WWW C WWW HTML HTML WWW WWW HTML HTML A B C 3 WWW 11),13),18) A B C A alt

4 82 Sep HTML B A A A WWW C A B HTML C WWW URL 3 C A B CBIR CBIR Image Collector WWW A CBIR B A A 2 A B ( 1) WWW A B (2) A 19) HTML C B A WWW 1 Fig. 1 Flow of gathering images from WWW. (3) B A A A A B A 20) 1 4. Image Collector HTML URL HTML URL WWW HTML

5 Vol. 42 No. SIG 10(TOD 11) WWW 83 URL HTML 2 Fig. 2 Overview of the gathering part. html file image file WWW HTML URL URL 3. URL URL 1 WWW HTML HTML URL 4. HTML HTML URL A B URL URL 5. URL HTML lion Web Web lion lion animal lion URL WWW HTML URL HTML URL AND URL Google Infoseek Lycos URL 1 HTTP Hypertext Transfer Protocol HTML URL HTML 100 URL URL URL 1 URL HTML URL 1 URL URL HTML URL HTML URL 4.1.4

6 84 Sep HTML URL URL URL HTML HTML URL HTML HTML HTML Web img src a href URL A B URL URL 1 URL HTML Web Html a href URL URL URL URL URL HTML 2 URL URL HTML HTML HTML URL HTML 3 (1) 3 HTML Fig. 3 HTML tags that are clues for the evaluation of the intensity of relation between an image and keywords. (2) img src alt 3(1) a href a href /a 3 (2) title meta name= "description" meta name="keyword" 3 (3) (H1,..,H6) 3 (4) 3 3 (5) 3 A 2 1 B 0 C A B A B HTML HTML WWW 11),13),18) 4.2 A B

Vol. 42 No. SIG 10(TOD 11) WWW 85 4.2.1 CBIR 21) RGB Lu v 22) 2 23) Lu v 6 6 6 6 216 216 4.2.2 A A 2 d ij 24) d ij =(h i h j ) t A (h i h j ) (1 ) A =[a ij ] (2) { 1 (i = j) a ij = e cd ij /d max (3) (i j) d max = max(d ij ) (4) ij h i h j i j c d ij 19) 0 255 RGB Lu v Lu v 6 Fig.

7 Vol. 42 No. SIG 10(TOD 11) WWW CBIR 21) RGB Lu v 22) 2 23) Lu v A A 2 d ij 24) d ij =(h i h j ) t A (h i h j ) (1 ) A =[a ij ] (2) { 1 (i = j) a ij = e cd ij /d max (3) (i j) d max = max(d ij ) (4) ij h i h j i j c d ij 19) RGB Lu v Lu v 6 Fig. 4 4 An example of a dendrogram of a clustering result. FN farthest neighbor method 4 m A A n C 1,.., C n C 1,.., C n C j m C j C j A A sel A sel = {a i A a i n j=1 C j} (5) B A B A A A B A A t

8 86 Sep A B A +B () % () %% Table 1 Experimental results. This table describes the number of collected images from WWW and the number of selected images from them. Numerical value in ( ) represents the precision and the recall. HTML A B (A +B ) 1, (85) 62 (94,95) 216 (26) 66 (42,49) 288 (41) 128 (67,73) 2, (86) 76 (95,87) 237 (50) 99 (72,60) 334 (60) 175 (82,71) 2, (48) 73 (53,95) 528 (74) 272 (83,58) 613 (70) 345 (77,62) 2, (90) 72 (92,97) 212 (50) 84 (71,56) 288 (60) 156 (81,72) 2, (95) 38 (95,97) 167 (60) 58 (73,43) 206 (67) 96 (82,57) 2, (71) 51 (75,95) 178 (33) 71 (42,50) 235 (42) 122 (56,69) 1, (95) 34 (97,92) 28 (25) 14 (36,72) 66 (65) 48 (79,88) 1, (71) 317 (91,75) 837 (42) 158 (66,30) 1,378 (53) 475 (82,53) B B B sel A A sel A sel B sel = {b i B min 1 j n d(b i, avg(c j ))<t} (6) A sel = {a i A sel min 1 j n d(a i, avg(c j ))<t} (7) d(a i,a j ) a i a j avg(c j ) C j A B A FN A 5. C Perl Linux AT Athlon 750 MHz 384 MB.jp WWW JPEG 1 8 Web URL Google Goo Infoseek Japan Lycos Japan OCN Navi Excite Japan 6 6 2,000 URL URL ,699 URL URL URL HTML 1,979 HTML 1,364 JPEG URL A 72 B HTML URL 1,699 HTML HTML HTML A 72 B 216 A 61 B 57 (A )/(A ) A 85% B A +B 26% 41%

9 Vol. 42 No. SIG 10(TOD 11) WWW 87 ( )/(WWW ) A 72 A 5% 9 60 A 5% (A )/(A ) (A )/(A ) 93% 92% (WWW ) (A ( B ) ) 5 A 2 B A A 9 B A 12 2 B 66 A 42% 49% A 2 2 A 94% 95% 6 B 2 5 A B (A ) 0.05 x x 3 5% % 73% 70 5 HTML URL1, % 1,364 ( ( ) )/( URL ) 86% 62% 1 8 HTML A B A +B A A 7 8 A B 1 8 A B A m m 1 10 m m m m =4 m

10 88 Sep A Fig. 5 Lion images selected from group A. 6 B Fig. 6 Lion images selected from group B. 7 A Fig. 7 Mt. Fuji images selected from group A. 8 B Fig. 8 Mt. Fuji images selected from group B. Fig. 9 9 Graphs of the precision and the recall of selected images and the precision of collected images.

11 Vol. 42 No. SIG 10(TOD 11) WWW 89 9 A A A lion 10 Fig. 10 The precision and the recall when changing the condition of selecting clusters. 11 B Fig. 11 The precision and the recall when changing the threshold of selecting images from B. B t 11 m =3 t t B 6. 9 A A 5% A 48% A % B A B 8 A A B

12 90 Sep WWW World-Wide Web Image Collector WWW B B A C 1 B CBIR 5 1) 2) 3) 4) Gudivada, V. and Raghavan, V.: Content- Based Image Retrieval-Systems, IEEE Comput., Vol.28, No.9, pp (1995). 5) Vol.40, No.SIG3(TOD 1), pp (1999). 6) Kiyoki, Y., Kitagawa, T. and Hayama, T.: A metadatabase system for semantic image search by a mathematical model of meaning, ACM SIGMOD Record, Vol.23, No.4, pp (1994). 7) Vol.33, No.11, pp (1992). 8) Weber, M., Welling, M. and Perona, P.: Towards Automatic Discovery of Object Categories, IEEE Computer Vision and Pattern Recognition, pp (2000). 9) Nelson, R. and Selinger, A.: Learning 3D Recognition Models for General Objects from Unlabeled Imagery: An Experiment in Intelligent Brute Force, International Conference on Pattern Recognition, Vol.I, pp.1 8 (2000). 10) WWW 15 Vol.15, No.3E1-05 (2001). 11) Framkel, C., Swain, M. and Athitsos, V.: Web- Seer: An Image Search Engine for the World Wide Web, Technical Report TR-96-14, University of Chicago (1996). 12) Smith, J. and Chang, S.: Visually Searching the Web for Content, IEEE Multimedia, Vol.4, No.3, pp (1997). 13) Sclaroff, S., LaCascia, M., Sethi, S. and Taycher, L.: Unifying Textual and Visual Cues for Content-Based Image Retrieval on the World Wide Web, Computer Vision and Image Understanding, Vol.75, No.1/2, pp (1999). 14) Rowley, H., Baluja, S. and Kanade, T.: Neural Network-Based Face Detection, IEEE Trans. Pattern Analysis and Machine Intelligence, Vol.20, No.1, pp (1998). 15) Smith, J. and Chang, S.: Visual SEEk: A Fully Automated Content-Based Image Query System, ACM International Conference on Multimedia 1996, pp (1996). 16) Bianchi-Berthoze, B. and Kato, T.: Towards a Comprehensive Integration of Subjective Parameters in Database Browsing, Advanced Database Systems for Integration of Media and User Environments 98, Vol.9, pp (1998). 17) Inder, N., Bianchi-Berthoze, B. and Kato, T.: K-DIME: A Software Framework for Kansei Filtering of Internet Material, IEEE International Conf. on Systems, Man and Cybernetics, Vol.6, pp (1999). 18) Rowe, N. and Frew, B.: Automatic caption localization for photographs on World-Wide

Vol. 42 No. SIG 10(TOD 11) WWW 91 Web pages, Information Processing and Management, Vol.34, No.1, pp.95 107 (1998). 19) (1986). 20) Vol.41, No.SIG3(TOD 5), pp.54 63 (2000). 21) Swain, M.

13 Vol. 42 No. SIG 10(TOD 11) WWW 91 Web pages, Information Processing and Management, Vol.34, No.1, pp (1998). 19) (1986). 20) Vol.41, No.SIG3(TOD 5), pp (2000). 21) Swain, M. and Ballard, D.: Color Indexing, International Journal of Computer Vision, Vol.7, No.1, pp (1991). 22) (1991). 23) (2000). 24) Hafner, J., Sawhney, H., Equitz, W., Flickner, M. and Niblack, W.: Efficient Color Histogram Indexing for Quadratic Form Distance Functions, IEEE Trans. Pattern Analysis and Machine Intelligence, Vol.17, No.7, pp (1995). ( ) ( ) WWW IEEE CS

Microsoft Word - toyoshima-deim2011.doc

Microsoft Word - toyoshima-deim2011.doc DEIM Forum 2011 E9-4 252-0882 5322 252-0882 5322 E-mail: t09651yt, sashiori, kiyoki @sfc.keio.ac.jp CBIR A Meaning Recognition System for Sign-Logo by Color-Shape-Based Similarity Computations for Images