Information: Join today and make your research activities more affordable! Technical workshop participation fees and annual registration fees are available at member rates.
Notice: [Important] Announcement of Changes to Registration Fee Payment and Manuscript Upload Procedures for IEICE Technical Meetings
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 2026-03-05 12:50
Low-Computational Flexible Deep Denoiser with Noise Scaling
Kon Kim, Makoto Nakashizuka (CIT) SIS2025-58
Abstract (in Japanese) (See Japanese page) 
(in English) Gaussian image denoising is a fundamental task in low-level vision, yet many deep learning--based approaches are trained under a fixed noise level and exhibit significant performance degradation when the noise condition differs at test time.
This limitation reduces their practical usability in real-world scenarios, where noise intensity often varies and cannot be precisely controlled. Diffusion-based deep denoisers provide an iterative refinement framework and show promising results, but they also suffer from limited generalization when the noise level is mismatched.
In this work, we propose a noise-adaptive diffusion-based deep denoising model designed to achieve robust performance across a wide range of Gaussian noise levels.
The key idea is to incorporate noise-level awareness into the iterative refinement process.
The proposed model introduces a scaling parameter of the denoising strength into each process of the iteration, allowing the correction magnitude to be dynamically adjusted according to the noise intensity.
This design enables the model to preserve fine image structures under low-noise conditions while providing stronger suppression for severe noise.
We evaluate the proposed scaling method on the widely used Set68 benchmark under multiple noise settings, including fixed noise levels ($sigma$ = 15, 25, 50) as well as randomly sampled noise conditions.
Experimental results demonstrate that conventional diffusion-based models achieve high performance only near their trained noise level, whereas the noise scaling method maintains stable and consistent PSNR performance across the entire noise range.
Visual comparisons further confirm that the proposed method reduces over-smoothing at low noise levels and better preserves structural details under high noise conditions.
These results indicate that the noise scaling is an effective strategy for improving robustness and generalization in image denoising.
Keyword (in Japanese) (See Japanese page) 
(in English) image denoising / Gaussian noise / diffusion process / deep learning / noise-level adaptation / iterative refinement / /  
Reference Info. IEICE Tech. Rep., vol. 125, no. 393, SIS2025-58, pp. 6-9, March 2026.
Paper # SIS2025-58 
Date of Issue 2026-02-26 (SIS) 
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 SIS2025-58

Conference Information
Committee SIS  
Conference Date 2026-03-05 - 2026-03-06 
Place (in Japanese) (See Japanese page) 
Place (in English) Ohkubo Campus, Saitama University 
Topics (in Japanese) (See Japanese page) 
Topics (in English)  
Paper Information
Registration To SIS 
Conference Code 2026-03-SIS 
Language English 
Title (in Japanese) (See Japanese page) 
Sub Title (in Japanese) (See Japanese page) 
Title (in English) Low-Computational Flexible Deep Denoiser with Noise Scaling 
Sub Title (in English)  
Keyword(1) image denoising  
Keyword(2) Gaussian noise  
Keyword(3) diffusion process  
Keyword(4) deep learning  
Keyword(5) noise-level adaptation  
Keyword(6) iterative refinement  
Keyword(7)  
Keyword(8)  
1st Author's Name Kon Kim  
1st Author's Affiliation Chiba Institute of Technology (CIT)
2nd Author's Name Makoto Nakashizuka  
2nd Author's Affiliation Chiba Institute of Technology (CIT)
3rd Author's Name  
3rd Author's Affiliation ()
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 ()
21st Author's Name  
21st Author's Affiliation ()
22nd Author's Name  
22nd Author's Affiliation ()
23rd Author's Name  
23rd Author's Affiliation ()
24th Author's Name  
24th Author's Affiliation ()
25th Author's Name  
25th Author's Affiliation ()
26th Author's Name / /
26th Author's Affiliation ()
()
27th Author's Name / /
27th Author's Affiliation ()
()
28th Author's Name / /
28th Author's Affiliation ()
()
29th Author's Name / /
29th Author's Affiliation ()
()
30th Author's Name / /
30th Author's Affiliation ()
()
31st Author's Name / /
31st Author's Affiliation ()
()
32nd Author's Name / /
32nd Author's Affiliation ()
()
33rd Author's Name / /
33rd Author's Affiliation ()
()
34th Author's Name / /
34th Author's Affiliation ()
()
35th Author's Name / /
35th Author's Affiliation ()
()
36th Author's Name / /
36th Author's Affiliation ()
()
Speaker Author-1 
Date Time 2026-03-05 12:50:00 
Presentation Time 20 minutes 
Registration for SIS 
Paper # SIS2025-58 
Volume (vol) vol.125 
Number (no) no.393 
Page pp.6-9 
#Pages 4 
Date of Issue 2026-02-26 (SIS) 

[Return to Top Page]

[Return to IEICE Web Page]
The Institute of Electronics, Information and Communication Engineers (IEICE), Japan