Unsupervised Seismic Data Denoising via SIREN‐Guided Diffusion Mode.
Saved in:
| Title: | Unsupervised Seismic Data Denoising via SIREN‐Guided Diffusion Mode. |
|---|---|
| Authors: | Li, Ji1 (AUTHOR), Liu, Dawei2 (AUTHOR) 409791715@qq.com, Trad, Daniel1 (AUTHOR) |
| Source: | Geophysical Prospecting. May2026, Vol. 74 Issue 4, p1-16. 16p. |
| Subject Terms: | *Signal denoising, *Implicit functions, *Machine learning, *Inverse problems, *Artificial neural networks, *Probabilistic generative models |
| Abstract: | Denoising diffusion probabilistic models have emerged as powerful generative frameworks capable of synthesizing high quality images through iterative denoising of Gaussian noise. However, their unconditional nature limits their effectiveness in inverse problems such as denoising, where reconstruction from a specific observation is required. Recent extensions like denoising diffusion restoration models introduce conditional generation by incorporating known degradation operators and pre‐trained denoisers, but these approaches remain heavily reliant on supervised training and large‐scale external datasets. This work proposes a fully unsupervised alternative called the sinusoidal representation network (SIREN)‐guided diffusion denoising model, which integrates implicit neural representations into the diffusion process. We replace the conventional pre‐trained denoiser with a coordinate‐based network (SIREN) optimized directly at inference time. Instead of assuming an explicit degradation model, we perform partial noise injection to the observed signal and initialize the reverse diffusion from this intermediate state. At each step, the SIREN is trained to match the current sample. This integration brings out the complementary strengths of both components: the diffusion process provides a structured, probabilistic coarse‐to‐fine regularization mechanism that stabilizes the optimization trajectory. At the same time, SIREN offers a powerful inductive bias for capturing structured, continuous signals. Crucially, this significantly eliminates the need for early stopping, a known challenge in SIREN‐based denoising, and makes the method highly robust to noise without supervision or clean training data. We evaluate our approach on synthetic and real seismic datasets and demonstrate that it achieves better denoising performance compared to traditional and deep learning‐based baselines. [ABSTRACT FROM AUTHOR] |
| Database: | Energy & Power Source |
| FullText | Text: Availability: 0 |
|---|---|
| Header | DbId: enr DbLabel: Energy & Power Source An: 194013175 AccessLevel: 6 PubType: Academic Journal PubTypeId: academicJournal PreciseRelevancyScore: 0 |
| IllustrationInfo | |
| Items | – Name: Title Label: Title Group: Ti Data: Unsupervised Seismic Data Denoising via SIREN‐Guided Diffusion Mode. – Name: Author Label: Authors Group: Au Data: <searchLink fieldCode="AR" term="%22Li%2C+Ji%22">Li, Ji</searchLink><relatesTo>1</relatesTo> (AUTHOR)<br /><searchLink fieldCode="AR" term="%22Liu%2C+Dawei%22">Liu, Dawei</searchLink><relatesTo>2</relatesTo> (AUTHOR)<i> 409791715@qq.com</i><br /><searchLink fieldCode="AR" term="%22Trad%2C+Daniel%22">Trad, Daniel</searchLink><relatesTo>1</relatesTo> (AUTHOR) – Name: TitleSource Label: Source Group: Src Data: <searchLink fieldCode="JN" term="%22Geophysical+Prospecting%22">Geophysical Prospecting</searchLink>. May2026, Vol. 74 Issue 4, p1-16. 16p. – Name: Subject Label: Subject Terms Group: Su Data: *<searchLink fieldCode="DE" term="%22Signal+denoising%22">Signal denoising</searchLink><br />*<searchLink fieldCode="DE" term="%22Implicit+functions%22">Implicit functions</searchLink><br />*<searchLink fieldCode="DE" term="%22Machine+learning%22">Machine learning</searchLink><br />*<searchLink fieldCode="DE" term="%22Inverse+problems%22">Inverse problems</searchLink><br />*<searchLink fieldCode="DE" term="%22Artificial+neural+networks%22">Artificial neural networks</searchLink><br />*<searchLink fieldCode="DE" term="%22Probabilistic+generative+models%22">Probabilistic generative models</searchLink> – Name: Abstract Label: Abstract Group: Ab Data: Denoising diffusion probabilistic models have emerged as powerful generative frameworks capable of synthesizing high quality images through iterative denoising of Gaussian noise. However, their unconditional nature limits their effectiveness in inverse problems such as denoising, where reconstruction from a specific observation is required. Recent extensions like denoising diffusion restoration models introduce conditional generation by incorporating known degradation operators and pre‐trained denoisers, but these approaches remain heavily reliant on supervised training and large‐scale external datasets. This work proposes a fully unsupervised alternative called the sinusoidal representation network (SIREN)‐guided diffusion denoising model, which integrates implicit neural representations into the diffusion process. We replace the conventional pre‐trained denoiser with a coordinate‐based network (SIREN) optimized directly at inference time. Instead of assuming an explicit degradation model, we perform partial noise injection to the observed signal and initialize the reverse diffusion from this intermediate state. At each step, the SIREN is trained to match the current sample. This integration brings out the complementary strengths of both components: the diffusion process provides a structured, probabilistic coarse‐to‐fine regularization mechanism that stabilizes the optimization trajectory. At the same time, SIREN offers a powerful inductive bias for capturing structured, continuous signals. Crucially, this significantly eliminates the need for early stopping, a known challenge in SIREN‐based denoising, and makes the method highly robust to noise without supervision or clean training data. We evaluate our approach on synthetic and real seismic datasets and demonstrate that it achieves better denoising performance compared to traditional and deep learning‐based baselines. [ABSTRACT FROM AUTHOR] |
| PLink | https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&db=enr&AN=194013175 |
| RecordInfo | BibRecord: BibEntity: Identifiers: – Type: doi Value: 10.1111/1365-2478.70172 Languages: – Code: eng Text: English PhysicalDescription: Pagination: PageCount: 16 StartPage: 1 Subjects: – SubjectFull: Signal denoising Type: general – SubjectFull: Implicit functions Type: general – SubjectFull: Machine learning Type: general – SubjectFull: Inverse problems Type: general – SubjectFull: Artificial neural networks Type: general – SubjectFull: Probabilistic generative models Type: general Titles: – TitleFull: Unsupervised Seismic Data Denoising via SIREN‐Guided Diffusion Mode. Type: main BibRelationships: HasContributorRelationships: – PersonEntity: Name: NameFull: Li, Ji – PersonEntity: Name: NameFull: Liu, Dawei – PersonEntity: Name: NameFull: Trad, Daniel IsPartOfRelationships: – BibEntity: Dates: – D: 01 M: 05 Text: May2026 Type: published Y: 2026 Identifiers: – Type: issn-print Value: 00168025 Numbering: – Type: volume Value: 74 – Type: issue Value: 4 Titles: – TitleFull: Geophysical Prospecting Type: main |
| ResultId | 1 |
