國立高雄大學圖資館 |

登入

回首頁

Deep Learning Approaches for Imaging Inverse Problems With Structured Noise /

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Deep Learning Approaches for Imaging Inverse Problems With Structured Noise /Jeffrey Alido.
作者:	Alido, Jeffrey,
面頁冊數:	1 electronic resource (121 pages)
附註:	Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
附註:	Advisors: Tian, Lei Committee members: Goyal, Vivek; Ohn-Bar, Eshed; Batmanghelich, Kayhan; Sun, Yu.
Contained By:	Dissertations Abstracts International87-03B.
標題:	Electrical engineering.
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=32239394
ISBN:	9798293806928

Deep Learning Approaches for Imaging Inverse Problems With Structured Noise /
Alido, Jeffrey,

Deep Learning Approaches for Imaging Inverse Problems With Structured Noise /Jeffrey Alido. - 1 electronic resource (121 pages)

Source: Dissertations Abstracts International, Volume: 87-03, Section: B.

Structured and spatially correlated noise presents a major challenge in scientific and biomedical imaging, where idealized assumptions of additive white Gaussian noise often break down. This dissertation addresses this challenge through two frameworks based on deep learning for solving inverse problems with structured noise, namely noise that is not additive white Gaussian noise (AWGN): a simulation-based supervised learning approach for low signal-to-background ratio (SBR) fluorescence imaging, and a novel generative modeling framework based on Whitened Score (WS) diffusion models for general imaging inverse problems with correlated Gaussian noise.The first part of this work introduces SBR-Net, a deep neural network trained on synthetic data generated by a structured background noise simulator that models light scattering and structured fluorescent background in thick biological tissue. This approach enables single-shot 3D volumetric reconstruction from light-field microscopy measurements with extremely low SBR. By explicitly modeling structured background noise and simulating realistic measurement-ground truth pairs, SBR-Net learns a direct inverse mapping. The framework is evaluated on synthetic and experimental data, with analysis of generalization behavior under real-world noise mismatch.The second part introduces Whitened Score (WS) diffusion models, a new class of generative priors tailored to inverse problems with structured noise. Conventional score-based diffusion models, trained on isotropic Gaussian noise, lack inductive biases suitable for real-world noise distributions encountered in applications such as diffraction tomography, interferometry, and wide-field microscopy. WS models reformulate the denoising objective by learning a whitened score function, where the standard (Stein) score vector field is whitened by the diffusion matrix of a stochastic differential equation, thus avoiding covariance inversion and enabling training under arbitrary Gaussian forward processes. This formulation allows WS models to serve as strong Bayesian priors, denoising structured noise and consistently outperforming conventional diffusion models across a range of computational imaging tasks.These contributions highlight the importance of aligning priors with real-world data and incorporating physical models and domain-specific noise characteristics to address inverse problems in realistic imaging settings. By combining simulation, supervised learning, and generative modeling, this work offers robust, interpretable solutions under structured and spatially correlated noise conditions.

English

ISBN: 9798293806928Subjects--Topical Terms:

454503
Electrical engineering.
Subjects--Index Terms:

Computational imaging

Deep Learning Approaches for Imaging Inverse Problems With Structured Noise /
LDR:04100nmm a22004333i 4500 001 693756
005 20260511063020.5
006 m o d
007 cr|nu||||||||
008 260819s2025 miu||||||m |||||||eng d
020 $a 9798293806928
035 $a (MiAaPQD)AAI32239394
035 $a AAI32239394
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Alido, Jeffrey, $e author. $0 (orcid)0000-0001-9730-5596 $3 1010416
245 1 0 $a Deep Learning Approaches for Imaging Inverse Problems With Structured Noise / $c Jeffrey Alido.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2025
300 $a 1 electronic resource (121 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
500 $a Advisors: Tian, Lei Committee members: Goyal, Vivek; Ohn-Bar, Eshed; Batmanghelich, Kayhan; Sun, Yu.
502 $b Ph.D. $c Boston University $d 2025.
520 $a Structured and spatially correlated noise presents a major challenge in scientific and biomedical imaging, where idealized assumptions of additive white Gaussian noise often break down. This dissertation addresses this challenge through two frameworks based on deep learning for solving inverse problems with structured noise, namely noise that is not additive white Gaussian noise (AWGN): a simulation-based supervised learning approach for low signal-to-background ratio (SBR) fluorescence imaging, and a novel generative modeling framework based on Whitened Score (WS) diffusion models for general imaging inverse problems with correlated Gaussian noise.The first part of this work introduces SBR-Net, a deep neural network trained on synthetic data generated by a structured background noise simulator that models light scattering and structured fluorescent background in thick biological tissue. This approach enables single-shot 3D volumetric reconstruction from light-field microscopy measurements with extremely low SBR. By explicitly modeling structured background noise and simulating realistic measurement-ground truth pairs, SBR-Net learns a direct inverse mapping. The framework is evaluated on synthetic and experimental data, with analysis of generalization behavior under real-world noise mismatch.The second part introduces Whitened Score (WS) diffusion models, a new class of generative priors tailored to inverse problems with structured noise. Conventional score-based diffusion models, trained on isotropic Gaussian noise, lack inductive biases suitable for real-world noise distributions encountered in applications such as diffraction tomography, interferometry, and wide-field microscopy. WS models reformulate the denoising objective by learning a whitened score function, where the standard (Stein) score vector field is whitened by the diffusion matrix of a stochastic differential equation, thus avoiding covariance inversion and enabling training under arbitrary Gaussian forward processes. This formulation allows WS models to serve as strong Bayesian priors, denoising structured noise and consistently outperforming conventional diffusion models across a range of computational imaging tasks.These contributions highlight the importance of aligning priors with real-world data and incorporating physical models and domain-specific noise characteristics to address inverse problems in realistic imaging settings. By combining simulation, supervised learning, and generative modeling, this work offers robust, interpretable solutions under structured and spatially correlated noise conditions.
546 $a English
590 $a School code: 0017
650 4 $a Electrical engineering. $3 454503
650 4 $a Optics. $3 204142
653 $a Computational imaging
653 $a Diffusion models
653 $a Generative modeling
653 $a Inverse problems
653 $a Optimization
690 $a 0544
690 $a 0800
690 $a 0752
710 2 $a Boston University. $b Electrical & Computer Engineering ENG. $3 915736
720 1 $a Tian, Lei $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 87-03B.
790 $a 0017
791 $a Ph.D.
792 $a 2025
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=32239394