IEICE Technical Committee Submission System
Conference Paper's Information
Online Proceedings
[Sign in]
Tech. Rep. Archives
 Go Top Page Go Previous   [Japanese] / [English] 

Paper Abstract and Keywords
Presentation 2023-03-17 15:40
Improvement of cross-attention modules for image captioning using pixel-wise semantic information
Zhihao Chen, Keisuke Doman, Yoshito Mekada (Chukyo Univ.) IMQ2022-84 IE2022-161 MVE2022-114
Abstract (in Japanese) (See Japanese page) 
(in English) Most of image captioning models have attention modules, and the module outputs an attention map (weighted feature map) from an image and a word sequence. However, a predicted word sequence may contain errors, and thus, the module output may not also be as expected. In this regard, we showed that pixel-wise semantic information (PSI) was effective for improving the attention modules in CNN-based models. This paper proposes the introduction of the PSI to cross attention modules in Transformer-based models. Experimental results showed that the PSI contributed the improvement of captioning accuracy on GRIT, a Transformer-based model. The accuracy of the proposed method, a BLEU4 score of 42.9, was equivalent to the sixth place on the COCO Captions Benchmark. This paper also studies grouping synonyms (e.g. ``table'' and ``coffee table'') in the PSI as a solution of the problem that attentions are distributed among synonyms, and discusses its effectiveness.
Keyword (in Japanese) (See Japanese page) 
(in English) Image Captioning / Vision and Language / Attention / Semantic segmentation / / / /  
Reference Info. IEICE Tech. Rep., vol. 122, no. 440, MVE2022-114, pp. 333-338, March 2023.
Paper # MVE2022-114 
Date of Issue 2023-03-08 (IMQ, IE, MVE) 
ISSN Online edition: ISSN 2432-6380
Copyright
and
reproduction
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034)
Download PDF IMQ2022-84 IE2022-161 MVE2022-114

Conference Information
Committee IMQ IE MVE CQ  
Conference Date 2023-03-15 - 2023-03-17 
Place (in Japanese) (See Japanese page) 
Place (in English) Okinawaken Seinenkaikan (Naha-shi) 
Topics (in Japanese) (See Japanese page) 
Topics (in English) Media of five senses, Multimedia, Media experience, Picture codinge, Image media quality, Network,quality and reliability, etc(AC) 
Paper Information
Registration To MVE 
Conference Code 2023-03-IMQ-IE-MVE-CQ 
Language Japanese 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Improvement of cross-attention modules for image captioning using pixel-wise semantic information 
Sub Title (in English)  
Keyword(1) Image Captioning  
Keyword(2) Vision and Language  
Keyword(3) Attention  
Keyword(4) Semantic segmentation  
Keyword(5)  
Keyword(6)  
Keyword(7)  
Keyword(8)  
1st Author's Name Zhihao Chen  
1st Author's Affiliation Chukyo University (Chukyo Univ.)
2nd Author's Name Keisuke Doman  
2nd Author's Affiliation Chukyo University (Chukyo Univ.)
3rd Author's Name Yoshito Mekada  
3rd Author's Affiliation Chukyo University (Chukyo Univ.)
4th Author's Name  
4th Author's Affiliation ()
5th Author's Name  
5th Author's Affiliation ()
6th Author's Name  
6th Author's Affiliation ()
7th Author's Name  
7th Author's Affiliation ()
8th Author's Name  
8th Author's Affiliation ()
9th Author's Name  
9th Author's Affiliation ()
10th Author's Name  
10th Author's Affiliation ()
11th Author's Name  
11th Author's Affiliation ()
12th Author's Name  
12th Author's Affiliation ()
13th Author's Name  
13th Author's Affiliation ()
14th Author's Name  
14th Author's Affiliation ()
15th Author's Name  
15th Author's Affiliation ()
16th Author's Name  
16th Author's Affiliation ()
17th Author's Name  
17th Author's Affiliation ()
18th Author's Name  
18th Author's Affiliation ()
19th Author's Name  
19th Author's Affiliation ()
20th Author's Name  
20th Author's Affiliation ()
Speaker Author-1 
Date Time 2023-03-17 15:40:00 
Presentation Time 25 minutes 
Registration for MVE 
Paper # IMQ2022-84, IE2022-161, MVE2022-114 
Volume (vol) vol.122 
Number (no) no.437(IMQ), no.439(IE), no.440(MVE) 
Page pp.333-338 
#Pages
Date of Issue 2023-03-08 (IMQ, IE, MVE) 


[Return to Top Page]

[Return to IEICE Web Page]


The Institute of Electronics, Information and Communication Engineers (IEICE), Japan