Presentation 2017-03-20
Calculation of Inter-Concept Distance based on Multi-Modal Topic Model Considering Junk Images
Yusuke Nagasawa, Kazuaki Nakamura, Naoko Nitta, Noboru Babaguchi,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) In this paper, we propose a method to calculate Inter-Concept Distance (ICD) using tagged images. In most cases, visual content of tagged images are correlated with their text tags, so it is not an appropriate strategy to handle visual features and textual features separately. However, there are some “junk images”, whose visual content and text tags are not related with each other. From these backgrounds, the proposed method jointly models the distributions of visual features and textual features for each concept by Multi-Modal Topic model (abbreviated as MMT model). In addition, to suppress the influence of junk images, the proposed method introduces the index named “image reliability”, which quantifies how likely the image is non-junk image, and iteratively performs the following two steps: (a) learning a MMT model for each concept based on the image reliability and (b) estimatingthe image reliability of each image based on the learned MMT models. In our experiments, the proposed MMT model can provide ICD closer to human perception than some other methods. Moreover, it is also indicated thatthe performance can be further improved by proposed iterative calculation of ICD and image reliability.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) Inter-Concept Distance / topic model / junk image / tagged image
Paper # BioX2016-50,PRMU2016-213
Date of Issue 2017-03-13 (BioX, PRMU)

Conference Information
Committee PRMU / BioX
Conference Date 2017/3/20(2days)
Place (in Japanese) (See Japanese page)
Place (in English)
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair Eisaku Maeda(NTT) / Masakatsu Nishigaki(Shizuoka Univ.)
Vice Chair Seiichi Uchida(Kyushu Univ.) / Hironobu Fujiyoshi(Chubu Univ.) / Akira Otsuka(AIST) / Hiroshi Takano(Toyama Pref. Univ.)
Secretary Seiichi Uchida(Kyoto Univ.) / Hironobu Fujiyoshi(NTT) / Akira Otsuka(NEC) / Hiroshi Takano(AIST)
Assistant Masaki Oonishi(AIST) / Takuya Funatomi(NAIST) / Masatsugu Ichino(Univ. of Electro-Comm.) / Naoyuki Takada(Secom) / Takahiro Aoki(Fujitsu Labs.)

Paper Information
Registration To Technical Committee on Pattern Recognition and Media Understanding / Technical Committee on Biometrics
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Calculation of Inter-Concept Distance based on Multi-Modal Topic Model Considering Junk Images
Sub Title (in English)
Keyword(1) Inter-Concept Distance
Keyword(2) topic model
Keyword(3) junk image
Keyword(4) tagged image
1st Author's Name Yusuke Nagasawa
1st Author's Affiliation Osaka University(Osaka Univ.)
2nd Author's Name Kazuaki Nakamura
2nd Author's Affiliation Osaka University(Osaka Univ.)
3rd Author's Name Naoko Nitta
3rd Author's Affiliation Osaka University(Osaka Univ.)
4th Author's Name Noboru Babaguchi
4th Author's Affiliation Osaka University(Osaka Univ.)
Date 2017-03-20
Paper # BioX2016-50,PRMU2016-213
Volume (vol) vol.116
Number (no) BioX-527,PRMU-528
Page pp.pp.105-110(BioX), pp.105-110(PRMU),
#Pages 6
Date of Issue 2017-03-13 (BioX, PRMU)