Diffusion-Models-for-Computational-Optical-Imaging
February 13, 2026 · View on GitHub
● As an emerging technology, computational optical imaging demonstrates remarkable advantages through its high degrees of freedom, high-dimensional information acquisition capability, and intelligent characteristics.
● The core challenge lies in achieving high-quality image reconstruction through inverse problem solving.
● While traditional regularization methods have improved reconstruction quality to some extent, their robustness and generalizability remain limited when handling complex imaging scenarios.
● Generative artificial intelligence offers a promising solution by learning data distributions to provide more universal and adaptive prior information, thereby enabling new possibilities for high-fidelity image reconstruction.
● Generative AI-assisted computational optical imaging [CITA2024 PPT]
● 面向物理属性复用的生成式智能计算光学成像 [PDF]
● 基于生成式人工智能的计算光学成像进展（特邀）[Paper]
● Learning from Spatial Domain
● The final output of computational optical imaging is typically represented in the spatial domain.
● Prior learning in spatial domain (including for grayscale images, color images, high-dimensional color images, and complex amplitude images), provides direct constraint generation method.
● Lens-less imaging via score-based generative model
● Multi-phase FZA Lensless Imaging via Diffusion Model
● Imaging through scattering media via generative diffusion model
● HoloDiffusion: Sparse Digital Holographic Reconstruction via Diffusion Modeling
● DMEDH: Diffusion Model-boosted Multiplane Extrapolation for Digital Holographic Reconstruction
● Fluorescence molecular tomography via score-based generative model
● High-resolution iterative reconstruction at extremely low sampling rate for Fourier single-pixel imaging via diffusion model
● Sparse-view reconstruction for photoacoustic tomography combining diffusion model with model-based iteration
● Score-based generative model-assisted information compensation for high-quality limited-view reconstruction in photoacoustic tomography
● Learning from Frequency Domain
● Computational optical imaging can transform signals into the frequency domain through hardware systems, where frequency-domain prior learning provides an additional dimension of constraints.
● Frequency domain generative diffusion model for temporal compressive coherent diffraction imaging
● Learning from Hybrid Domain
● Some computational optical imaging systems employ iterative spatial/frequency domain cycling.
● Joint spatial-frequency learning can further enhance imaging quality.
● Dual-domain mean-reverting diffusion model-enhanced temporal compressive coherent diffraction imaging
● Learning from Transformed Domain
● Learning in transform domains (e.g., wavelet domain) can effectively enhance image reconstruction details in computational optical imaging.
● Wavelet-refinement-inspired diffusion model for scattering imaging
[paper] [code]
● Learning from Data Domain
● For sparse-sampled computational imaging problems, learning directly in the native data domain of signals provides a more straightforward generative sparse reconstruction scheme.
● Multiple diffusion models-enhanced extremely limited-view reconstruction strategy for photoacoustic tomography boosted by multi-scale priors
● Ultra-sparse reconstruction for photoacoustic tomography: sinogram domain prior-guided method exploiting enhanced score-based diffusion model
● Other related work
● Real-time intelligent 3D holographic photography for real-world scenarios
● High-speed real 3D scene acquisition and 3D holographic reconstruction system based on ultrafast optical axial scanning
[paper]
● Low-cost, high-precision integral 3D photography and holographic 3D display for real-world scenes
[paper]