saguar1

YANAI Lab.

電気通信大学 総合情報学科/大学院 総合情報学専攻 メディア情報学コース 柳井研究室
電気通信大学 > 情報工学科 > コンピュータ学講座 > 柳井研究室 > 研究紹介  

?Á´¶¸ì¤ËÂ?±þ¤?¤ë?èÁü¤Îǧ?±?ÄÇ?À?¤ÎÉ??Á

??ÅÄ¡¡ÏÂ


Date: Ê¿À® 27 ǯ 2 ?î 17 Æü


1 ¤Ï¤¸¤á¤Ë

¤?¤ì¤Þ¤Ç¡¤°ìÈÌŪ¤ÊʪÂÎ?èÁü¡¤¿©?ö?èÁü¡¤ÁǺà?èÁü¤Îǧ?±¤Ê¤É¡¤ÍÍ¡?¤ÊʪÂÎǧ?±¤Î¸¦?椬?Ô¤ï¤ì¤Æ¤?¤¿¡? Ëܸ¦?æ¤Ç¤Ï¡¤ÊªÂΤÎ?Á´¶¤ËÃåÌÜ¡¤ºÇ¿?¤Î?èÁüǧ?±?êË¡¤òÍѤ¤¤Æ¡¤?Á´¶¤òÉ?¸?¤?¤ë¸ÀÍÕ¤ËÂ?±þ¤?¤ë?èÁü¤Îǧ?±?ÄÇ?À?¤ÎÉ??Á¤ò?Ô¤¦¡? ?Á´¶¤ÏʪÂΤÎ?Á¤Ë°Í¸¤?¤Ê¤¤¤Î¤Çǧ?±¤¬Æñ¤?¤¤¤?¤È¤¬Í?¬¤?¤ì¤ë¤¬¡¤¶áǯ¡¤?èÁüǧ?±¤ÎÀºÅ٤ϸþ?å¤?¤Æ¤ª¤ê¡¤?èÁü¤?¤éʪÂΤÎ?Á´¶¤òǧ?±¤Ç¤?¤ë?ÄÇ?À?¤¬¤¢¤ë¡?

Ëܸ¦?æ¤Ç¤Ï¡¤?Á´¶¤òÉ?¸?¤?¤ë¸ì¤È¤?¤Æ?ª?Î?Þ?È?Ú¤òÍѤ¤¤¿¡? ?ª?Î?Þ?È?Ú¤ËÂ?±þ¤?¤ëWeb?èÁü¤ò?ý?¸¤?¡¤Ç§?±?ÄÇ?À?¤òÉ??Á¤?¤ë¡?

2 ´ØÏ¢¸¦?æ

´ØÏ¢¸¦?æ¤È¤?¤Æ¤Ï¡¤Cimpoi[1]¤é¤Ê¤É¤ÎÁǺà?èÁüǧ?±¤Î¸¦?椬¤¢¤ë¡? Liu¤é¤Ë¤è¤ëFlickr Material Database(FMD)[2]¤Îǧ?±¤¬ÂåÉ?Ū¤Ç¤¢¤ë¡? FMD¤ÏÁ¡°Ý, ?¬?é??¡¤¶â°¤Ê¤É¤Î10?ïÎà¤ÎÁǺà???Æ?´?ê¤?¤é¤Ê¤ë?Ç¡??¿???Ã?Ȥǡ¤¤?¤ì¤é¤Î?èÁü¤òʬÎà¤?¤ë¸¦?椬À?¤ó¤Ë¤?¤ì¤Æ¤?¤¿¡?

3 ??Í×

Ëܸ¦?æ¤Ç¤Ï¡¤Web?èÁü¤Ë¤ª¤±¤ë?Î?¤?º?èÁü¤ò??Æ°¤Ç?ü?î¤?¡¤?ª?Î?Þ?È?Ú¤ËÂ?±þ¤?¤ë?èÁü¤ò?ý?¸¤?¤¿¡? ¿Þ1¤Ë?èÁü¤ò?ý?¸¤?¤ëºÝ¤Îή¤ì¤ò?¨¤?¤¿¡?
Figure 1: ?Ç¡??¿???Ã?È??ÃÛ¤Îή¤ì

4 ?êË¡¤Î?ܺÙ

1 ?èÁü¤Î?ý?¸

Bing Image search API¤òÍѤ¤¤ÆWeb?èÁü¤Î?ý?¸¤ò?Ԥä¿¡? ?Æ?¯?¨?ê¤Ë¤Ä¤¤¤ÆAPI¤Î?å¸Â¤Î1000Ëç¤Þ¤Ç?èÁü¤ò?ý?¸¤?¤¿¡?

2 ?Î?¤?º?èÁü¤Î?ü?î

WebAPI¤Ç?ý?¸¤?¤¿?èÁü¤Ë¤Ï¡¤?å°Ì¤Î?èÁü¤Ç¤¢¤Ã¤Æ¤â?Î?¤?º?èÁü¤¬´Þ¤Þ¤ì¤Æ¤?¤Þ¤¦¤?¤È¤¬¤¢¤ë¡? ¤?¤?¤Ç¡¤?èÁüǧ?±¤Ë¤è¤ë?ê?é?ó???ó?°¤Ë¤è¤ê?Î?¤?º?èÁü¤ò?ü?î¤?¤ë¡?

Ëܸ¦?æ¤Ç¤Ï¡¤???÷?Ý?¸?Æ?£?Ö?èÁü¤È¤?¤Æ¸¡º÷?ë?Ì?å°Ì¤Î?èÁü¡¤?Í?¬?Æ?£?Ö???ó?×?ë¤Ë?é?ó?À?à?èÁü¤òÍѤ¤¤ÆSVM¤ò?Ø?¬¤?¤ë¡? SVM¤ò?Æ?èÁü¤ËŬÍѤ?¤¿ºÝ¤ËÆÀ¤é¤ì¤ë??ÎÏÃͤÎÂç?®¤Ç¡¤?èÁü¤Ë?é?ó???ó?°¤òÍ¿¤¨¡¤?ê?é?ó???ó?°¤ò?Ô¤¦¡?

3 ǧ?±?ÄÇ?À?¤ÎÉ??Á

À??ò?èÁü¤ò50 Ë硤ÉÔÀ??ò?èÁü¤ò5000 Ëç¤È¤?¤Æ¡¤¤?¤ì¤ò?Ø?¬¤?SVM?â?Ç?ë¤òºî¤ë¡? ¤?¤ÎSVM?â?Ç?ë¤ò?Ȥäơ¤5050Ëç¤Î?èÁü¤òʬÎ?¤?¡¤¤?¤ÎʬÎ?ÅÙ?礤¤Ç¡¤Ç§?±?ÄÇ?À?¤òÉ??Á¤?¤¿¡? ¶ñÂÎŪ¤ÊÀºÅÙ¡¤¿ôÃͤϡ¤Ê¿¶ÑŬ?çΨ¤òÍѤ¤¤¿¡?

4 ?èÁüÆÃħÎÌ

?èÁüÆÃħÎ̤ÎÉ?¸?¤ÏImproved Fisher Vector ¤È Deep Convolutional Neural Network (DCNN) ¤òÍѤ¤¡¤?±ÊÌ´ï¤Ë¤ÏÀþ?ÁSVM¤òÍѤ¤¤ë¡?

1 Improved Fisher Vector¡ÊIFV¡Ë[3]

Fisher Vector¤Ïº®?ç?¬?¦??ʬÉÛ¤òÍøÍѤ?¤¿???Õ?ÈÎÌ?Ò??¤Ë¤è¤ê¡¤ÆÃħÎ̤ò?¨?ó??¡??ɤ?¤ë?êË¡¤Ç¤¢¤ë¡? Ëܸ¦?æ¤Ç¤Ï¡¤128?¡¸?¤ÎSURFÆÃħÎ̤ò1000¸Ä?é?ó?À?à???ó?×?ê?ó?°¤?¡¤?¨?ó??¡??ɤ?¤¿¡? ?¬?¦??ʬÉÛ¤Î?¯?é???¿¿ô¤Ï256¤È¤?¤¿¡?

2 Deep Convolutional Neural Network (DCNN)[4]

Ëܸ¦?æ¤Ç¤Ï¡¤?ª¡??×?ó??¡???¤Ç¤¢¤ëOverfeat[5]¤òÍøÍѤ?¤ÆDCNNÆÃħ¤òÃê??¤?¤Æ¤¤¤ë¡? Overfeat¤Ï¡¤ImageNet Challenge¤Î1000???Æ?´?ê¤Çpre-training¤?¤ì¤Æ¤¤¤ë¡? Ëܸ¦?æ¤Ç¤Ï¡¤Convolutional¤ÎºÇ¸å¤ÎÁؤÎLayer5¤È¡¤fully-connectedÁؤÎLayer6¤ÈLayer7¤Î??ÎÏ?ë?̤òL2À??¬??¤?ÆÃħÎ̤Ȥ?¤Æ°?¤¦¡?

3 Support Vector Machine¡ÊSVM¡Ë

?±Ê̤ˤÏSVM ¤òÍѤ¤¤ë¡? IFV¡¤DCNN¤Ï?â?¡¸?¤Ê¤Î¤ÇÀþ?ÁSVM ¤òÍøÍѤ?¤ë¡?

5 ?¸?

Æó???ïÎà¤Î?ª?Î?Þ?È?Ú¤òÍѤ¤¤Æ¡¤?Á´¶?èÁü?Ç¡??¿???Ã?ȤÎ??ÃÛ¡¤?Á´¶?èÁü¤Îǧ?±Î¨¤ÎÉ??Á¤ò?Ԥä¿

1 ?Ç¡??¿???Ã?ȤÎ??ÃÛ

É?1¤Ë?Ç¡??¿???Ã?ȤÎÀºÅÙ¤ÎÊ¿¶ÑÃͤò?¨¤?¤¿¡?

Table 1: ?Ç¡??¿???Ã?ȤÎÀºÅÙ¤ÎÊ¿¶ÑÃÍ
ÆÃħÎÌ Fisher DCNN L5 DCNN L6 DCNN L7
Ê¿¶ÑÃÍ 60.4% 85.2% 89.6% 85.1%

DCNNÆÃħ¤Î?ë?̤ÏFisher vector¤ÈÈæ?Ó¤?¤ë¤È¤É¤ÎLayer¤Ë¤ª¤¤¤Æ¤â¤è¤¤?ë?̤Ȥʤ俤¬¡¤ÆäËlayer6¤Ë¤è¤ë?ê?é?ó???ó?°¤ÎÀºÅÙ¤¬¤è¤?¤Ã¤¿¡?

¿Þ2¤ËDCNNÆÃħ¤ÎLayer6¤Ë¤è¤ë20?ïÎà¤Î?ª?Î?Þ?È?Ú?èÁü?Ç¡??¿???Ã?ȤÎÎã¤ò?¨¤?¡?

Figure 2: ?Ç¡??¿???Ã?ȤÎÎã

2 ǧ?±?ÄÇ?À?¤ÎÉ??Á

ǧ?±?ÄÇ?À?¤ò?é?ó?À?à?èÁü¤È¤ÎʬÎ?ÅÙ¤ÇÉ??Á¤?¤¿¡? ¤Þ¤¿¡¤É??Á¤Ï5 fold cross varidation¤Ç?Ԥä¿¡? É?2¤ËÉ??Á¤ÎÊ¿¶ÑÃÍ¡¤¿Þ3¤Ë?Æ?ª?Î?Þ?È?Ú¤ÎÉ??Á?ë?̤ò?¨¤?¤¿¡?¡?


Table 2: ǧ?±?ÄÇ?À?¤ÎÊ¿¶ÑÃÍ
ÆÃħÎÌ Fisher DCNN L5 DCNN L6 DCNN L7
Ê¿¶ÑÃÍ 21.5% 55.6% 51.3% 73.9%

Figure 3: ǧ?±?ë?Ì

3 Ì??ì¡Ü?ª?Î?Þ?È?Ú¤ÎʬÎà

?ª?Î?Þ?È?Ú¤Ç?ý?¸¤?¤¿?èÁü¤ÏÆÃÄê¤ÎʪÂΤËÊ?¤ë¤?¤È¤¬¤¢¤ê¡¤°Ọ̃Ū¤Êǧ?±¤ò?Ô¤¨¤Æ¤¤¤ë¤?ÉÔ?Τ?¤Ê?ì?礬¤¢¤Ã¤¿¡? ¤?¤?¤Ç¡¤Æ±°ìÌ??ìÆâ¤Ç¤Î?ª?Î?Þ?È?Ú?Þ?ë?Á?¯?é??ʬÎà¤ò?Ԥä¿¡? ¸¤¡¤?±¡???¡¤?¤¡¤?Ö¤Î?ͤĤÎÌ??ì¤ÇÌ??ì¡Ü?ª?Î?Þ?È?Ú?èÁü¤ò?ý?¸¤?¡¤Æ±°ìÌ??ìÆâ¤Ç?ª?Î?Þ?È?Ú?Þ?ë?Á?¯?é??ʬÎà¤ò?Ԥä¿¡?

1 Ì??ì¤ËÂ?±þ¤?¤ë?ª?Î?Þ?È?Ú¤Î?èÄê

Web?Æ?????ȸ¡º÷¤òÍøÍѤ?¤Æ¡¤Ì??ì¤Î??¡??Ñ??¤òºîÀ®¤?¡¤¶¦?¯ÉÑÅÙ¤Ç?ª?Î?Þ?È?Ú¤òÁªÂò¤?¤¿¡? º£?ó¤Ï¶¦?¯ÉÑÅÙ¤ÇÆÀ¤é¤ì¤ë?ª?Î?Þ?È?Ú¤¬?¯¤Ê¤?¤Ã¤¿¤Î¤Ç¡¤?ÁÍÆ?ì¤âÍѤ¤¤¿¡?

2 Ʊ°ìÌ??ìÆâ?ª?Î?Þ?È?Ú?Þ?ë?Á?¯?é??ʬÎà

¿Þ4¤Ë?±¡???¤Ë¤ª¤±¤ë?ª?Î?Þ?È?Ú?Þ?ë?Á?¯?é??¤ÎʬÎà?ë?̤ò?¨¤?¤¿¡?
Figure 4: ?±¡???¡Ü?ª?Î?Þ?È?Ú?èÁü¤ÎʬÎà?ë?Ì
?Þ?ë?Á?¯?é??¤ÎʬÎàΨ¡Ê7?¯?é??¡Ë¤Ï¡¤72.3%¡¤ ?ª?Î?Þ?È?Ú?¯?é??¤Î¤ß¤ÎʬÎàΨ¤Ï¡¤61.5%¤È¤Ê¤Ã¤¿¡?

6 ?Í?¡

?ª?Î?Þ?È?Ú?èÁü¤È?é?ó?À?à?èÁü¤ÎʬÎ?¤Ï¡¤?ë?̤¬?é?ó?À?à¤Ç¤¢¤Ã¤¿?ì?硤0.05%ÄøÅ٤ˤʤë¤?¤È¤ò?ͤ¨¤ë¤È¡¤Îɤ¤ÀºÅÙ¤ÇʬÎà¤?¤ë¤?¤È¤¬¤Ç¤?¤¿¡? ¤?¤?¤?¡¤ÊªÂÎǧ?±¤Î???¤?¢??¤Ë¤è¤ê¡¤Ç§?±¤¬ÍưפˤʤäƤ¤¤ë?ÄÇ?À?¤¬¤¢¤Ã¤¿¡?

¤?¤?¤Ç¡¤ÊªÂÎǧ?±¤Î???¤?¢??¤òÍÞ¤¨¤ë¤¿¤á¤Ë¡¤Æ±°ìÌ??ìÆâ¤Ç?Þ?ë?Á?¯?é??ʬÎà¤ò?Ԥä¿¡? 6¡Á8?¯?é??¡Ê300¡Á400Ëç¡Ë¤Ç¤ÎʬÎà¤Ç¤¢¤Ã¤¿¤?¤È¤ò?Íθ¤?¤ë¤È¤è¤¤ÀºÅÙ¤ÇʬÎà¤?¤ë¤?¤È¤¬¤Ç¤?¤¿¡? ?Á´¶¤Îǧ?±¤ËÂ?¤?¤ÆDCNNÆÃħ¤ÏIFV¤ÈÈæ¤Ù¤ÆÍ?¸ú¤Ç¤¢¤ë¤?¤È¤¬¸À¤¨¤ë¡?

7 º£¸å¤Î?ÝÂê

°Ọ̃Ū¤Êǧ?±¤ÎÀºÅÙ¤òÀ?¤?¤¯É??Á¤?¤ë¤¿¤á¤Ë¡¤ÊªÂÎǧ?±¤Î???¤?¢??¤òÍÞ¤¨¤ëɬÍפ¬¤¢¤ë¡? fine-tuning¤ä?Æ?ì?¤?ä¡?¤Î?Ø?¬ÊýË¡¤ÎÊÑ??¤Ê¤É¤Ë¤è¤ê¡¤CNN¤ò?Á´¶¤Îǧ?±¤ËºÇŬ??¤?¤ë¤?¤È¤Ç?þÁ±¤¬´üÂԤǤ?¤ë¡?

???Íʸ¸?

1
M. Cimpoi, S. Maji, I Kokkinos, S. Mohamed, and A. Vedaldi.
Describing textures in the wild.
In Proc. of IEEE Computer Vision and Pattern Recognition, 2014.

2
C. Liu, L. Sharan, E. Adelson, and R. Rosenholtz.
Exploring features in a bayesian framework for material recognition.
In Proc. of IEEE Computer Vision and Pattern Recognition, 2010.

3
F. Perronnin, J. Sanchez, and T. Mensink.
Improving the fisher kernel for large-scale image classification.
In Proc. of European Conference on Computer Vision, 2010.

4
A. Krizhevsky, I. Sutskever, and G E. Hinton.
Imagenet classification with deep convolutional neural networks.
In Advances in Neural Information Processing Systems, 2012.

5
S. Pierre, E. David, Z. Xiang, M. Michael, F. Rob, and L Yann.
Overfeat: Integrated recognition, localization and detection using convolutional networks.
In Proc. of International Conference on Learning Representations, 2014.