Presentation 2020-03-05
Self-Play Reinforcement Learning for Fast Image Retargeting
Nobukatsu Kajiura, Satoshi Kosugi, Xueting Wang, Toshihiko Yamasaki, Kiyoharu Aizawa,
PDF Download Page PDF download Page Link
Abstract(in Japanese) (See Japanese page)
Abstract(in English) We address image retargeting, which is a task of adjusting input images into arbitrary sizes. In a previous method, they combine multiple operators and generate multiple retargeted images to find the optimal combination, which minimizes the distance between the original and the retargeted images. In this paper, to find the optimal combination more quickly, we propose a method of predicting the optimal operator step by step using a reinforcement learning agent. The advantage of this method is that it is hardly affected even if the number of operators increases. Since the distance between the input image and the retargeted image varies greatly depending on the image, it is difficult to use the evaluation function of the conventional method as a reward. In order to solve this problem, we propose that a reward based on self-play can be insensitive to changes in the value of the evaluation function. We conduct experiments, which show that our method achieves multi-operator image retargeting that is faster by four orders of magnitude and has the same performance as the previous method.
Keyword(in Japanese) (See Japanese page)
Keyword(in English) image retargeting / multi-operator / deep reinforcement learning / self-play
Paper # IMQ2019-40,IE2019-122,MVE2019-61
Date of Issue 2020-02-27 (IMQ, IE, MVE)

Conference Information
Committee IE / IMQ / MVE / CQ
Conference Date 2020/3/5(2days)
Place (in Japanese) (See Japanese page)
Place (in English) Kyushu Institute of Technology
Topics (in Japanese) (See Japanese page)
Topics (in English)
Chair Hideaki Kimata(NTT) / Toshiya Nakaguchi(Chiba Univ.) / Kenji Mase(Nagoya Univ.) / Hideyuki Shimonishi(NEC)
Vice Chair Kazuya Kodama(NII) / Keita Takahashi(Nagoya Univ.) / Mitsuru Maeda(Canon) / Kenya Uomori(Osaka Univ.) / Masayuki Ihara(NTT) / Jun Okamoto(NTT) / Takefumi Hiraguri(Nippon Inst. of Tech.)
Secretary Kazuya Kodama(NTT) / Keita Takahashi(NHK) / Mitsuru Maeda(Shizuoka Univ.) / Kenya Uomori(Sony Semiconductor Solutions) / Masayuki Ihara(Nagoya Univ.) / Jun Okamoto(NTT) / Takefumi Hiraguri(Nippon Inst. of Tech.)
Assistant Kyohei Unno(KDDI Research) / Norishige Fukushima(Nagoya Inst. of Tech.) / Hiroaki Kudo(Nagoya Univ.) / Masaru Tsuchida(NTT) / Keita Hirai(Chiba Univ.) / Satoshi Nishiguchi(Oosaka Inst. of Tech.) / Masanori Yokoyama(NTT) / Shogo Fukushima(Univ. of ToKyo) / Chikara Sasaki(KDDI Research) / Yoshiaki Nishikawa(NEC) / Takuto Kimura(NTT)

Paper Information
Registration To Technical Committee on Image Engineering / Technical Committee on Image Media Quality / Technical Committee on Media Experience and Virtual Environment / Technical Committee on Communication Quality
Language JPN
Title (in Japanese) (See Japanese page)
Sub Title (in Japanese) (See Japanese page)
Title (in English) Self-Play Reinforcement Learning for Fast Image Retargeting
Sub Title (in English)
Keyword(1) image retargeting
Keyword(2) multi-operator
Keyword(3) deep reinforcement learning
Keyword(4) self-play
1st Author's Name Nobukatsu Kajiura
1st Author's Affiliation The University of Tokyo(UTokyo)
2nd Author's Name Satoshi Kosugi
2nd Author's Affiliation The University of Tokyo(UTokyo)
3rd Author's Name Xueting Wang
3rd Author's Affiliation The University of Tokyo(UTokyo)
4th Author's Name Toshihiko Yamasaki
4th Author's Affiliation The University of Tokyo(UTokyo)
5th Author's Name Kiyoharu Aizawa
5th Author's Affiliation The University of Tokyo(UTokyo)
Date 2020-03-05
Paper # IMQ2019-40,IE2019-122,MVE2019-61
Volume (vol) vol.119
Number (no) IMQ-454,IE-456,MVE-457
Page pp.pp.127-131(IMQ), pp.127-131(IE), pp.127-131(MVE),
#Pages 5
Date of Issue 2020-02-27 (IMQ, IE, MVE)