Freezing pre-trained parameters of encoders for denoisers: Expanding pixel involvement and filtering out high-frequency noise.

Bibliographic Details
Title: Freezing pre-trained parameters of encoders for denoisers: Expanding pixel involvement and filtering out high-frequency noise.
Authors: Zhang J; Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: jzhang151@connect.hkust-gz.edu.cn., Li J; Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: jli842@connect.hkust-gz.edu.cn., Feng L; Wave Functional Metamaterial Research Facility, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: liangfeng@hkust-gz.edu.cn., Zhang B; Department of Chemistry, The University of Hong Kong (Pokfulam), Hong Kong, 999077, China. Electronic address: bzhangay@connect.ust.hk., Luo X; Department of Physics, The Hong Kong University of Science and Technology, Hong Kong, 999077, China. Electronic address: xluoay@connect.ust.hk., Wen W; Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: phwen@ust.hk.
Source: Neural networks : the official journal of the International Neural Network Society [Neural Netw] 2026 Aug; Vol. 200, pp. 108794. Date of Electronic Publication: 2026 Mar 01.
Publication Type: Journal Article
Language: English
Journal Info: Publisher: Pergamon Press Country of Publication: United States NLM ID: 8805018 Publication Model: Print-Electronic Cited Medium: Internet ISSN: 1879-2782 (Electronic) Linking ISSN: 08936080 NLM ISO Abbreviation: Neural Netw Subsets: MEDLINE
Imprint Name(s): Original Publication: New York : Pergamon Press, [c1988-
MeSH Terms: Image Processing, Computer-Assisted*/methods , Deep Learning* , Neural Networks, Computer*, Signal-To-Noise Ratio ; Algorithms ; Humans
Abstract: Image denoising is a pivotal and intricate task within the field of deep learning and the challenge of generalization poses significant difficulties for many denoisers when faced with out-of-distribution (OOD) noise. Model generalization performance depends critically on three components: model architectural design; dataset composition; and carefully designed training methods. While the first two factors have consistently been the focus of research, there remains a significant gap in the exploration of training methods. In this paper, we prove a straightforward and efficient training strategy to improve the generalization of denoisers. Specifically, we propose that freezing the encoders' pre-trained parameters and updating the decoders' parameters during the training process. Experimental results confirm that our proposed method, despite its simplicity, achieves more robust denoising performance across various noise types. This demonstrates the effectiveness of our training strategy in enhancing network generalization. To the best of our knowledge, we are the first to demonstrate that freezing pre-trained parameters of encoders during the training process not only expands the range of input pixels that strongly influence the denoising results but also effectively filters out high-frequency noise signals, thereby improving the performance of the denoiser.
(Copyright © 2026 Elsevier Ltd. All rights reserved.)
Competing Interests: Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: JIE ZHANG reports was provided by The Hong Kong University of Science and Technology(GZ). If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Contributed Indexing: Keywords: Freezing parameters; Image denoising; Image processing; Network generalization
Entry Date(s): Date Created: 20260309 Date Completed: 20260710 Latest Revision: 20260710
Update Code: 20260711
DOI: 10.1016/j.neunet.2026.108794
PMID: 41797192
Database: MEDLINE
FullText Links:
  – Type: other
    Url: https://resolver.ebsco.com:443/public/rma-ftfapi/ejs/direct?AccessToken=4E359AB0B3231D2C6A0F&Show=Object
Text:
  Availability: 0
CustomLinks:
  – Url: https://www.doi.org/10.1016/j.neunet.2026.108794?
    Name: ScienceDirect (all content) (s7799221)
    Category: fullText
    Text: View record from ScienceDirect
    MouseOverText: View record from ScienceDirect
Header DbId: cmedm
DbLabel: MEDLINE
An: 41797192
AccessLevel: 3
PubType: Academic Journal
PubTypeId: academicJournal
PreciseRelevancyScore: 0
IllustrationInfo
Items – Name: Title
  Label: Title
  Group: Ti
  Data: Freezing pre-trained parameters of encoders for denoisers: Expanding pixel involvement and filtering out high-frequency noise.
– Name: Author
  Label: Authors
  Group: Au
  Data: <searchLink fieldCode="AU" term="%22Zhang+J%22">Zhang J</searchLink>; Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: jzhang151@connect.hkust-gz.edu.cn.<br /><searchLink fieldCode="AU" term="%22Li+J%22">Li J</searchLink>; Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: jli842@connect.hkust-gz.edu.cn.<br /><searchLink fieldCode="AU" term="%22Feng+L%22">Feng L</searchLink>; Wave Functional Metamaterial Research Facility, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: liangfeng@hkust-gz.edu.cn.<br /><searchLink fieldCode="AU" term="%22Zhang+B%22">Zhang B</searchLink>; Department of Chemistry, The University of Hong Kong (Pokfulam), Hong Kong, 999077, China. Electronic address: bzhangay@connect.ust.hk.<br /><searchLink fieldCode="AU" term="%22Luo+X%22">Luo X</searchLink>; Department of Physics, The Hong Kong University of Science and Technology, Hong Kong, 999077, China. Electronic address: xluoay@connect.ust.hk.<br /><searchLink fieldCode="AU" term="%22Wen+W%22">Wen W</searchLink>; Advanced Materials Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 51140, China. Electronic address: phwen@ust.hk.
– Name: TitleSource
  Label: Source
  Group: Src
  Data: <searchLink fieldCode="JN" term="%228805018%22">Neural networks : the official journal of the International Neural Network Society</searchLink> [Neural Netw] 2026 Aug; Vol. 200, pp. 108794. <i>Date of Electronic Publication: </i>2026 Mar 01.
– Name: TypePub
  Label: Publication Type
  Group: TypPub
  Data: Journal Article
– Name: Language
  Label: Language
  Group: Lang
  Data: English
– Name: TitleSource
  Label: Journal Info
  Group: Src
  Data: <i>Publisher: </i><searchLink fieldCode="PB" term="%22Pergamon+Press%22">Pergamon Press </searchLink><i>Country of Publication: </i>United States <i>NLM ID: </i>8805018 <i>Publication Model: </i>Print-Electronic <i>Cited Medium: </i>Internet <i>ISSN: </i>1879-2782 (Electronic) <i>Linking ISSN: </i><searchLink fieldCode="IS" term="%2208936080%22">08936080 </searchLink><i>NLM ISO Abbreviation: </i>Neural Netw <i>Subsets: </i>MEDLINE
– Name: PublisherInfo
  Label: Imprint Name(s)
  Group: PubInfo
  Data: <i>Original Publication</i>: New York : Pergamon Press, [c1988-
– Name: SubjectMESH
  Label: MeSH Terms
  Group: Su
  Data: <searchLink fieldCode="MM" term="%22Image+Processing%2C+Computer-Assisted%22">Image Processing, Computer-Assisted*</searchLink>/<searchLink fieldCode="MM" term="%22Image+Processing%2C+Computer-Assisted+methods%22">methods</searchLink> <br /><searchLink fieldCode="MM" term="%22Deep+Learning%22">Deep Learning*</searchLink> <br /><searchLink fieldCode="MM" term="%22Neural+Networks%2C+Computer%22">Neural Networks, Computer*</searchLink><br /><searchLink fieldCode="MH" term="%22Signal-To-Noise+Ratio%22">Signal-To-Noise Ratio</searchLink> ; <searchLink fieldCode="MH" term="%22Algorithms%22">Algorithms</searchLink> ; <searchLink fieldCode="MH" term="%22Humans%22">Humans</searchLink>
– Name: Abstract
  Label: Abstract
  Group: Ab
  Data: Image denoising is a pivotal and intricate task within the field of deep learning and the challenge of generalization poses significant difficulties for many denoisers when faced with out-of-distribution (OOD) noise. Model generalization performance depends critically on three components: model architectural design; dataset composition; and carefully designed training methods. While the first two factors have consistently been the focus of research, there remains a significant gap in the exploration of training methods. In this paper, we prove a straightforward and efficient training strategy to improve the generalization of denoisers. Specifically, we propose that freezing the encoders' pre-trained parameters and updating the decoders' parameters during the training process. Experimental results confirm that our proposed method, despite its simplicity, achieves more robust denoising performance across various noise types. This demonstrates the effectiveness of our training strategy in enhancing network generalization. To the best of our knowledge, we are the first to demonstrate that freezing pre-trained parameters of encoders during the training process not only expands the range of input pixels that strongly influence the denoising results but also effectively filters out high-frequency noise signals, thereby improving the performance of the denoiser.<br /> (Copyright © 2026 Elsevier Ltd. All rights reserved.)
– Name: Abstract
  Label: Competing Interests
  Group: Ab
  Data: Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: JIE ZHANG reports was provided by The Hong Kong University of Science and Technology(GZ). If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
– Name: SubjectMinor
  Label: Contributed Indexing
  Group:
  Data: <i>Keywords: </i>Freezing parameters; Image denoising; Image processing; Network generalization
– Name: DateEntry
  Label: Entry Date(s)
  Group: Date
  Data: <i>Date Created: </i>20260309 <i>Date Completed: </i>20260710 <i>Latest Revision: </i>20260710
– Name: DateUpdate
  Label: Update Code
  Group: Date
  Data: 20260711
– Name: DOI
  Label: DOI
  Group: ID
  Data: 10.1016/j.neunet.2026.108794
– Name: AN
  Label: PMID
  Group: ID
  Data: 41797192
PLink https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&db=cmedm&AN=41797192
RecordInfo BibRecord:
  BibEntity:
    Identifiers:
      – Type: doi
        Value: 10.1016/j.neunet.2026.108794
    Languages:
      – Code: eng
        Text: English
    PhysicalDescription:
      Pagination:
        StartPage: 108794
    Subjects:
      – SubjectFull: Signal-To-Noise Ratio
        Type: general
      – SubjectFull: Algorithms
        Type: general
      – SubjectFull: Humans
        Type: general
      – SubjectFull: Image Processing, Computer-Assisted methods
        Type: general
      – SubjectFull: Deep Learning
        Type: general
      – SubjectFull: Neural Networks, Computer
        Type: general
    Titles:
      – TitleFull: Freezing pre-trained parameters of encoders for denoisers: Expanding pixel involvement and filtering out high-frequency noise.
        Type: main
  BibRelationships:
    HasContributorRelationships:
      – PersonEntity:
          Name:
            NameFull: Zhang J
      – PersonEntity:
          Name:
            NameFull: Li J
      – PersonEntity:
          Name:
            NameFull: Feng L
      – PersonEntity:
          Name:
            NameFull: Zhang B
      – PersonEntity:
          Name:
            NameFull: Luo X
      – PersonEntity:
          Name:
            NameFull: Wen W
    IsPartOfRelationships:
      – BibEntity:
          Dates:
            – D: 01
              M: 08
              Text: 2026 Aug
              Type: published
              Y: 2026
          Identifiers:
            – Type: issn-electronic
              Value: 1879-2782
          Numbering:
            – Type: volume
              Value: 200
          Titles:
            – TitleFull: Neural networks : the official journal of the International Neural Network Society
              Type: main
ResultId 1