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