Random Structured Illumination Imaging

• Random structured illumination is a family of imaging strategies that uses nonuniform, often random or pseudo‐random, patterns to encode spatial information for inverse reconstruction.
• It improves imaging performance by enhancing localization, super-resolution, and robustness across modalities such as fluorescence microscopy, ghost imaging, and X-ray scattering.
• Reconstruction leverages techniques like linear inversion, compressed sensing, and deep learning to recover high-quality images even under low signal-to-noise conditions.

Taken together, the recent literature suggests that random structured illumination is best understood as a family of sensing and imaging strategies in which nonuniform illumination patterns supply measurement diversity for inverse reconstruction. The randomness may reside in the illumination field itself, as in random phase masks, random speckle, or pseudo-random projected dots; it may arise from randomized relative translations between a fixed pattern and the object; or it may be replaced by deterministic or learned codes that serve the same encoding role. Across microscopy, ghost imaging, phase retrieval, active stereo, hyperspectral sensing, grazing-incidence X-ray scattering, and free-electron-laser imaging, the common objective is to turn illumination into a sensing code that improves localization, super-resolution, robustness, or identifiability beyond what uniform illumination permits (Fannjiang, 2010).

1. Definition, scope, and neighboring concepts

In the strictest sense, random structured illumination uses a known or statistically characterized nonuniform pattern whose realization varies across measurements. In compressed-sensing-style sparse imaging, this appears as independent random probes implemented by random phase modulation; in phase retrieval, the measured object is the modulated field $\tilde f(\mathbf n)=f(\mathbf n)\lambda(\mathbf n)$, with $\lambda(\mathbf n)$ a known random mask; and in fluorescence microscopy it commonly appears as speckle excitation patterns, either known, estimated, or treated through their covariance (Fannjiang, 2011).

The literature also includes several closely related regimes that are not identical to direct random pattern projection. In Speckle Flow SIM, a single fixed speckle pattern is kept throughout acquisition and sample motion supplies the effective pattern diversity; the same conceptual shift appears in Stochastically Structured Illumination Microscopy (S$^2$IM), where stochastic object motion through a fixed speckle field replaces scanner-controlled phase stepping (Cao et al., 5 Feb 2025). In blind translated-pattern SIM, the illumination can remain fixed in the laboratory frame while the sample is translated, so the ensemble of illumination-object interactions becomes effectively randomized through the translation sequence (Capalbo et al., 24 Mar 2025).

Two adjacent ideas delimit the concept. First, several works show that structured illumination need not be random: grazing-incidence X-ray scattering with a 256-bit de Bruijn sequence coded aperture is explicitly deterministic, and the diversity arises from controlled scanning rather than stochasticity (Gursoy et al., 7 May 2025). Second, active stereo demonstrates that pseudo-random patterns are only a heuristic baseline: a learned diffractive optical element and reconstruction network can outperform fixed pseudo-random dot patterns, so “random-looking” illumination is not inherently optimal (Baek et al., 2020).

2. Forward models and physical encoding mechanisms

A generic computational-imaging formulation is sequential structured illumination with a bucket detector. For a thin semi-transparent object with projected transmission $X(\mathbf r)$, illuminated by patterns $I_m(\mathbf r)$, the bucket measurement is

$$a_m = \eta \int_{\Omega} X(\mathbf r)\, I_m(\mathbf r)\, d\mathbf r,$$

and reconstruction is a linear synthesis from the measured coefficients. In this model, random illumination is a special case of arbitrary linearly independent illumination patterns; the reconstructed image is the projection of the object onto the span of the chosen patterns (Gureyev et al., 2018).

In fluorescence SIM and blind SIM, the standard forward model is multiplicative illumination followed by diffraction-limited detection:

$$y_m = \mathcal{H} \otimes (\rho \times I_m) + \varepsilon_m,$$

or, for one SIM subframe,

$H(p) = (I p * \mathrm{PSF}) + N.$

These formulations support both known-pattern and unknown-pattern reconstructions. They also make clear why random speckle excitation is useful: the illumination multiplies the fluorophore distribution before low-pass filtering by the microscope PSF, thereby mixing object frequencies into the passband (Labouesse et al., 2016).

Motion-driven variants retain the same multiplicative physics but change the source of diversity. In Speckle Flow SIM, a translated sample obeys

$$D(\mathbf{r}) = \left(I(\mathbf{r}) \cdot S(\mathbf{r}+\delta\mathbf{r})\right) * \mathrm{PSF}(\mathbf{r}),$$

while in S$^2$IM the motion-corrected model is

$\lambda(\mathbf n)$0

with each frame seeing the same illumination shifted by a frame-dependent displacement (Fusco et al., 2024).

Other modalities instantiate the same encoding principle with different physics. In active stereo, the projector is modeled by wave optics and the cameras by geometric optics: the diffractive optical element induces phase delay, the modulated complex field is Fourier propagated, and the projected intensity pattern $\lambda(\mathbf n)$1 is then warped into each camera view using disparity and occlusion masks (Baek et al., 2020). In single-shot free-electron-laser imaging, the exit wave is $\lambda(\mathbf n)$2, and the detector records

$\lambda(\mathbf n)$3

where the known randomized probe $\lambda(\mathbf n)$4 supplies the coding needed for single-shot inversion (Levitan et al., 27 Feb 2026).

3. Resolution, sparsity, orthogonality, and uniqueness

A general resolution law for structured-illumination computational imaging is that, under the paper’s two stated conditions, the squared spatial resolution satisfies

$\lambda(\mathbf n)$5

This formalizes the intuition that $\lambda(\mathbf n)$6 linearly independent illuminations over area $\lambda(\mathbf n)$7 define roughly $\lambda(\mathbf n)$8 effective resolution elements. The same analysis shows that orthogonality is decisive for noise performance: with photon shot noise, orthogonal patterns attain the intuitive Poisson limit, whereas non-orthogonal patterns introduce reconstruction-induced spatial correlations and therefore lower SNR (Gureyev et al., 2018).

Random illumination becomes especially powerful when combined with sparsity priors. For sparse inverse imaging under the Born approximation, independent random probes make the sensing matrix sufficiently incoherent that OST, Lasso, and BPDN admit recovery guarantees. The principal scaling is

$\lambda(\mathbf n)$9

up to logarithmic factors, where $^2$0 is sparsity and $^2$1 is the number of measurements. With sufficiently many random probes, the Lasso can also beat the Rayleigh limit and localize subwavelength structure (Fannjiang, 2010).

A different theoretical role of random illumination appears in phase retrieval. Multiplying the object by a known random mask changes the algebraic structure of the object’s $^2$2-transform. For complex-valued objects of full-rank support, the modulated object is almost surely irreducible, which removes the nontrivial factorization ambiguities that ordinarily plague Fourier-magnitude phase retrieval. Under mild additional constraints, uniqueness up to global phase holds with probability one, and with two independent illuminations almost sure uniqueness up to global phase is established for general complex-valued objects without any constraint (Fannjiang, 2011).

In blind SIM, the decisive mechanism is not randomness alone but sparsity of the illumination-induced product images together with positivity. The reformulation $^2$3 converts blind SIM into a sequence of positivity-constrained deconvolutions. The paper shows numerically that standard speckle shifted upward by a positive constant behaves almost like wide-field deconvolution, whereas thresholded, sparser speckle that produces many zero regions yields clear super-resolution. This reinterprets random structured illumination in SIM as a “near-black” inverse problem in the sense of illumination-induced sparsity (Labouesse et al., 2016).

4. Reconstruction strategies

The reconstruction methodology depends on whether the illumination is known, unknown, estimated jointly, or represented statistically. In generic structured-illumination systems, linear least-squares, pseudoinverse inversion, and nonnegative least squares remain fundamental. Grazing-incidence X-ray structured illumination for GI diffraction uses the linear model $^2$4 and reconstructs localized scattering profiles with nonnegative least squares; random LED hyperspectral illumination uses the measurement model $^2$5, analyzes $^2$6 by SVD, and reconstructs with the Moore–Penrose pseudoinverse, truncating singular values below the noise floor (Gursoy et al., 7 May 2025).

Sparse and blind formulations broaden this picture. The compressed-sensing random-illumination framework studies OST, Lasso, and BPDN for exact localization and stable recovery (Fannjiang, 2010). Blind SIM recasts each frame as a positivity-constrained sparse deconvolution with regularizer $^2$7, solved efficiently by a preconditioned primal-dual splitting method rather than by slower constrained least squares (Labouesse et al., 2016). PE-SIMS estimates the unknown patterns from the data, assuming the sum of the patterns is uniform, then reconstructs from the covariance image $^2$8, with an additional pixel reassignment step in SIMS-PR to improve high-frequency OTF behavior (Yeh et al., 2016).

Learning-based methods fall into two distinct classes. The first is physics-informed optimization without training data: an untrained U-Net-like architecture is optimized directly on one measured SIM stack through the known forward model, using SSIM loss between generated and measured subframes (Burns et al., 2022). The second class is joint sensing-and-reconstruction optimization. In active stereo, the diffractive optical element phase delay $^2$9 and reconstruction network parameters $X(\mathbf r)$0 are learned end-to-end through a supervised disparity loss; the network is explicitly trinocular, fusing a wide stereo baseline with a narrow baseline between the camera and the illumination module (Baek et al., 2020).

Motion-based random structured illumination requires joint inference over object and motion. Speckle Flow SIM uses a neural space-time model (NSTM) to estimate both the motion field and the super-resolved scene from raw fluorescence frames under a fixed pre-calibrated speckle pattern (Cao et al., 5 Feb 2025). S$X(\mathbf r)$1IM alternates gradient-descent updates of the object and effective illumination, using the Kullback–Leibler divergence for low-photon experimental data (Fusco et al., 2024). Blind translated-pattern SIM uses a generalized Richardson–Lucy algorithm that alternates latent-frame updates with averaged updates of the object and unknown pattern; the paper reports that randomized translations are less artifact-prone than ordered grids (Capalbo et al., 24 Mar 2025).

5. Experimental realizations across imaging modalities

In fluorescence microscopy, random structured illumination has developed along several lines. An untrained PINN for SIM reconstructs linear SIM, nonlinear SIM, and localized plasmonic SIM by merely changing the known illumination patterns in the loss function. On simulated data, the reported metrics are PSNR around 32–35 dB, SSIM around 0.90–0.94, and low NRMSE; the reported resolution improvements are about 2× diffraction limit for linear SIM and about 3× diffraction limit for nonlinear SIM and localized plasmonic SIM (Burns et al., 2022). Speckle Flow SIM shows that 10 raw images under one fixed speckle pattern can outperform 40 raw images in speckle SIM with unknown pattern estimation, and it was demonstrated experimentally on 0.71 µm fluorescent microbeads (Cao et al., 5 Feb 2025).

S$X(\mathbf r)$2IM translates the same principle to scan-less ophthalmoscopic super-resolution. Using a motorized phantom eye and stochastic eye-motion statistics, the reported test-target resolution improves from 6.5 ± 0.2 μm in wide field to 3.4 ± 0.1 μm in S$X(\mathbf r)$3IM, corresponding to $X(\mathbf r)$4 and a headline enhancement of 1.91× (Fusco et al., 2024). A hardware-oriented branch of the field addresses illumination generation itself: a zwitterion-doped liquid-crystal dynamic speckle generator produces statistically independent high-contrast speckles with tunable decorrelation time from 0.1 s to 0.1 ms, enabling 2 µm axial resolution for optical sectioning and 1.5-fold lateral resolution improvement in widefield random illumination fluorescence microscopy (Magermans et al., 27 Feb 2026).

Blind-SIM variants emphasize robustness rather than explicit pattern control. In the generalized Richardson–Lucy formulation, numerical results give a resolution enhancement of 2.076 for C-SIM and 1.94 for Gen-RL in nearly noiseless data, while experimental datasets M1 and M2 show that Gen-RL is more robust in the short-exposure noisy regime; the paper states that about 20 iterations are sufficient and that randomized translations reduce artifacts relative to ordered ones (Capalbo et al., 24 Mar 2025).

Beyond fluorescence microscopy, random structured illumination spans several distinct application areas. In active stereo, a fabricated DOE implementing the learned Polka Lines pattern achieves a mean absolute depth error of 1.4 cm over 0.4–1.0 m on textureless planar targets and is reported to be more robust than the Intel RealSense D415 pattern in high-dynamic-range scenes (Baek et al., 2020). In ghost imaging, sinusoidally modulated structured speckle illumination yields a factor-of-2 sub-Rayleigh improvement, from 433 $X(\mathbf r)$5m to 216.5 $X(\mathbf r)$6m, and can be combined with pseudo-inverse ghost imaging (Li, 2024). In grazing-incidence X-ray scattering, the deterministic coded-aperture analogue uses a 15 keV beam at 0.3°, a 256-bit de Bruijn-sequence aperture scanned over 600 $X(\mathbf r)$7m in 1 $X(\mathbf r)$8m steps, and reports about 190 $X(\mathbf r)$9m resolution along the z-direction for localized scattering reconstruction (Gursoy et al., 7 May 2025).

At shorter wavelengths, randomized illumination enables single-shot inversion. Randomized probe imaging at a free-electron laser uses a randomized zone plate to generate a speckled XUV probe and reports robust single-shot reconstructions at 400 nm full pitch over a 40 μm diameter field of view; for resonant magnetic-domain imaging, the reported single-shot resolution is about 2.22 µm full pitch, improving to the pixel-limited 1.04 µm full pitch after averaging (Levitan et al., 27 Feb 2026). A spectral rather than spatial realization appears in LED hyperspectral sensing, where random subsets of discrete LEDs with random intensity form the illumination matrix; for sparse vegetation spectra, reconstruction reaches RMSE less than 1% with as few as 25 LEDs under 1% noise (Howell et al., 27 Oct 2025).

6. Limitations, misconceptions, and conceptual boundaries

A common misconception is that “random” automatically implies “better.” The general computational-imaging analysis shows that non-orthogonal random patterns can degrade SNR because the inversion step creates spatial correlations in the reconstructed image; orthogonality remains the favorable case for shot-noise-limited performance (Gureyev et al., 2018). Active stereo provides a complementary caution: pseudo-random dot fields are widely used, but jointly learned illumination can outperform both pseudo-random and regular dot baselines under equal illumination power, which indicates that randomness alone is not the relevant design criterion (Baek et al., 2020).

Another important distinction is between known, unknown, and statistically known illuminations. The untrained PINN framework assumes that both pattern shape and pattern location are known; reconstruction quality degrades if SNR is too low, modulation depth is poor, or the patterns are not sufficiently informative (Burns et al., 2022). PE-SIMS relaxes exact pattern knowledge but assumes that the average illumination is uniform and that the pattern covariance is known or estimable; it also requires enough frames for stable variance estimation, otherwise shading artifacts remain (Yeh et al., 2016). Blind generalized Richardson–Lucy methods dispense with exact pattern calibration, but their success depends on stable translations and avoidance of over-iteration (Capalbo et al., 24 Mar 2025).

Motion-driven random structured illumination has its own constraints. Speckle Flow SIM requires informative motion, temporal redundancy, and negligible intra-exposure blur; if parts of the sample do not move, the inverse problem becomes ill-posed, and abrupt dynamics can violate the NSTM model (Cao et al., 5 Feb 2025). S$I_m(\mathbf r)$0IM similarly requires accurate motion estimation and rigid-translation approximations over short exposures; its experimental implementation therefore used reflectance-based registration with accuracy below 0.125 pixel and temporally rearranged frames so successive registered views were close in displacement (Fusco et al., 2024).

Finally, the boundary between random structured illumination and deterministic coded illumination is methodologically significant. The de Bruijn coded aperture for surface-resolved grazing-incidence X-ray scattering is explicitly not random, and its reconstruction diversity comes from designed coding plus scanning rather than from stochasticity (Gursoy et al., 7 May 2025). This suggests that the central issue is often not randomness itself but whether the illumination acts as a sufficiently informative sensing code for the inverse problem at hand. In some regimes that code is best implemented by random masks or speckle; in others it is better implemented by deterministic codes, learned optical elements, or randomized relative motion.