---
title: Optics-Informed Intensity Adjustment (OIA)
url: https://www.emergentmind.com/topics/optics-informed-intensity-adjustment-oia
type: topic
---

# Optics-Informed Intensity Adjustment (OIA)

Optics-informed Intensity Adjustment (OIA) denotes a family of optical methods in which intensity is treated as an explicit object of control, calibration, inversion, or learned correction, rather than as a secondary consequence of phase manipulation alone. In published work from 2019 to 2025, the label encompasses pupil-plane intensity correction in adaptive optics, spectrally calibrated microscopy correction, intensity-corrected light-in-flight reconstruction, AI-driven contrast tuning in OCT-SLO, universal linear intensity transformations under spatially-incoherent illumination, longitudinal focal shaping in high-intensity optics, and optics-driven neural feature adjustment for metalens endoscopy. Taken together, these works suggest a common principle: measured or modelled optical priors are used to compensate spatial non-uniformity, total energy error, spectral imbalance, or application-specific intensity distortions [2405.16289] [2103.12464] [1903.06519] [2408.02703] [2303.13037] [2106.06802] [2509.23294] [2508.03596].

## 1. Conceptual scope and recurring structure

Across the literature, OIA does not denote a single algorithm. Rather, it appears as a recurring strategy in which the optical system is endowed with an intensity model, an intensity sensor or proxy metric, and a compensator or inversion rule. In some cases, the compensator is a pixelated intensity corrector in a pupil-conjugate plane; in others it is a per-pixel calibration curve, a diffractive processor, a variable-focus liquid lens and reference-arm sweep, an aspheric mirror sag profile, or a learned attention module. The shared feature is that the correction is informed by optical quantities such as intensity distribution, total energy, spectral throughput, Rayleigh scattering, focusing geometry, point-spread-function structure, or wavelength-dependent efficiency.

| Setting | Adjusted quantity | Representative mechanism |
|---|---|---|
| Adaptive optics | non-uniform intensity distribution and net energy loss at the pupil plane | pixelated SLM with crossed polarizers and dual-feedback loop |
| Light-in-flight imaging | optics- and atmosphere-modulated SPAD intensity | forward model and per-pixel correction factor |
| Bright-field microscopy | raw counts to photon-proportional flux | spectral transfer function and per-pixel calibration curves |
| OCT-SLO | sample specific automatic contrast adjustment of the beam | AI scoring over liquid-lens voltage and reference-arm length |
| Diffractive processing | arbitrary linear transformation in time-averaged intensity | phase-only diffractive layers under incoherent illumination |
| Metalens endoscopy | intensity decay and radial vignetting | optical embeddings and channel/spatial attention |

This breadth is a source of both utility and terminological ambiguity. A common misconception is to treat OIA as synonymous with amplitude flattening in adaptive optics. The published record is broader: some OIA methods are closed-loop control schemes, some are inverse physical models, some are fabrication-time optical designs, and some are end-to-end learned modules embedded in larger computational imaging systems.

## 2. Pupil-plane intensity control in adaptive optics

The most explicit formalization of OIA appears in "Intensity adaptive optics" [2405.16289], which introduces intensity adaptive optics (I-AO) as an adaptive-optics framework that corrects not only phase and polarisation but also spatial and total energy errors in an optical system. I-AO treats the pupil-plane intensity map as a controlled degree of freedom by inserting, in conjugate to the system pupil, a pixelated intensity corrector implemented as an SLM between crossed polarizers, together with a two-stage feedback controller. The first loop addresses non-uniform intensity distribution; the second compensates for energy loss at the pupil plane.

The intensity-error model is written as
$$
I_{\mathrm{err}}(x,y)=T(x,y)\,I_{\mathrm{in}}(x,y), \qquad 0 \le T(x,y) \le 1.
$$
Using the SLM as a controllable retarder, the pixel transmittance is
$$
T_{\mathrm{SLM}}(x,y)=\sin^2[\delta(x,y)/2].
$$
The corrected distribution is chosen so that
$$
I_{\mathrm{corr}}(x,y)=T(x,y)\,T_{\mathrm{SLM}}(x,y)\,I_{\mathrm{in}}(x,y)\simeq I_{\mathrm{ref}}(x,y),
$$
which yields the pixel-wise retardance pattern
$$
\delta^*(x,y)=2\,\arcsin\!\left[\sqrt{I_{\mathrm{ref}}(x,y)/I_{\mathrm{err}}(x,y)}\right].
$$
After the first loop, the residual net energy error is
$$
\Delta E=\int_\Omega [I_{\mathrm{ref}}(x,y)-I_{\mathrm{corr}}(x,y)]\,dx\,dy,
$$
and the second loop sets the attenuator according to
$$
A^*=E_{\mathrm{ref}}/E_{\mathrm{corr}}\equiv \int_\Omega I_{\mathrm{ref}} \big/ \int_\Omega I_{\mathrm{corr}}.
$$

Two implementation pathways are given. In the sensor-based pathway, a pupil-conjugate camera measures $I_{\mathrm{err}}(x,y)$, the SLM applies $\delta^*(x,y)$, and a deformable mirror then runs a standard phase AO routine to remove residual $\phi(x,y)$. In the sensorless pathway, the system sequentially applies intensity Zernike modes $Z_n(x,y)$ with trial amplitudes $a_n$, evaluates focal-plane image quality, and maximizes
$$
J(a)=\alpha\,C_{\mathrm{low}}(a)+\beta\,C_{\mathrm{circ}}(a), \qquad \alpha+\beta=1,
$$
before a final DM-based phase correction.

Performance is quantified by the pupil-plane uniformity metric
$$
U=1-\sigma_I/\mu_I
$$
and the Strehl ratio
$$
S=I_{\mathrm{peak}}/I_{\mathrm{Airy}}.
$$
Representative results are reported as follows: without I-AO, $U \approx 0.60$ and $S \approx 0.35$; sensor-based I-AO after SB1 yields $U \approx 0.94$, $S \approx 0.78$, and after SB2 yields $U \approx 0.98$, $S \approx 0.85$; sensorless I-AO after SL1 yields $U \approx 0.90$, $S \approx 0.72$, and after SL2 yields $U \approx 0.96$, $S \approx 0.80$. These improvements, described as an $\sim 50$–$70\%$ increase in $U$ and $>2\times$ in $S$, are used to argue that conventional phase-only AO cannot recover either spatial uniformity or absolute energy when intensity errors dominate.

## 3. Physics-based inversion and calibration workflows

A second major strand of OIA uses explicit forward models and calibrated transfer functions to demodulate measured intensities. In "Intensity-corrected 4D light-in-flight imaging" [2103.12464], the recorded SPAD intensity for a laser pulse propagating in air is modeled as
$$
I(x;\theta)=B\cdot I_f(x;\theta)\cdot I_r(x;\theta)\cdot I_s(x;\theta),
$$
where the factors account for geometric or optical focusing, Rayleigh scattering, and finite integration length along the path subtended by one pixel. The inversion fits the measured amplitudes $A_i$ and arrival times $t_i$ to the intensity model and time-delay model, computes a per-pixel correction factor
$$
F_i=I_{\mathrm{model}}(x_i;\theta^*)/C^*,
$$
and recovers the true intensity as
$$
I_{\mathrm{true}}(x_i)=A_i/F_i.
$$
For pulses traveling toward the camera at nominal $\theta=167^\circ$, the joint fit returned $\theta^*=165.5^\circ\pm0.1^\circ$ with residual RMS error on the intensity map $<4\%$; for pulses traveling away at $\theta=13^\circ$, the fit returned $\theta^*=13.9^\circ\pm0.1^\circ$ with intensity residuals again $<5\%$. The apparent camera-space velocities were superluminal or subluminal depending on geometry, but after correction the pulse propagates at $c$ in real time. The central-pixel intensity versus angle matched
$$
I_c(\theta)\propto (1+\cos^2\theta)/\sin\theta
$$
with $R^2>0.98$ and typical point-wise error $<3\%$.

In bright-field transmission microscopy, "Spectroscopic Approach to Correction and Visualisation of Bright-Field Light Transmission Microscopy Biological Data" [1903.06519] formulates OIA as a spectrally informed calibration of the entire optical path. The channel-specific transfer function is defined as
$$
H_c(\lambda)=QE_c(\lambda)
$$
or, when independently measured path transmission is included,
$$
H_c(\lambda)=T_{\mathrm{opt}}(\lambda)\,QE_c(\lambda).
$$
The photon-proportional radiant flux reaching a pixel of channel $c$ behind filter $k$ is
$$
\Phi_{k,c}=\int_0^\infty L_k(\lambda)\,H_c(\lambda)\,d\lambda,
$$
and per-pixel calibration curves $f_{x,y,c}: s\to \phi$ convert dark-corrected raw counts to corrected intensities
$$
I_{\mathrm{corr}}(x,y,c)=f_{x,y,c}\bigl(s_0(x,y,c)\bigr).
$$
The corrected data are then compressed to 8 bpc using the Least Information Loss algorithm, whose objective is described as preserving every occupied gray level and thereby minimizing information loss.

The AI-driven OCT-SLO formulation is different again. In "Self-calibrating Intelligent OCT-SLO System" [2408.02703], OIA is realized as sample specific automatic contrast adjustment of the beam on a pre-instructed region of interest. A variable-focus liquid lens and a stepper-motorized OCT reference arm are swept over settings; for each setting the system acquires 360 B-scans and constructs a 3D OCT volume and an en-face projection; an AI observer scores each volume and locks in the voltage and mirror-position settings with the highest score. The final implementation uses MobileNet-V2, which runs in $\approx 0.15$ s per volume versus $\approx 2$–$3$ s for U-Net. Relative to AI-driven automation, a purely SNR-maximization loop was reported to be $\sim 200\%$ less accurate and $\sim 130\%$ slower, while a complete two-knob sweep and convergence occurred in $\sim 55$ s versus $>10$ min manually. The reported spatial resolution results were 2.41 $\mu$m in a glass-bead phantom in the axial direction, 0.76 with STD 0.46 $\mu$m in mouse retina in the axial direction, and better than 228 line pair per millimeter or 2 $\mu$m for all three spectrums in the $x$–$y$ plane.

## 4. Diffractive processors and learned optics-driven adjustment

Under spatially-incoherent illumination, OIA can be implemented as a universal linear map in intensity. "Universal Linear Intensity Transformations Using Spatially-Incoherent Diffractive Processors" [2303.13037] shows that the time-averaged output intensity can be written as
$$
O(m,n)=\sum_{m',n'} |h(m,n;m',n')|^2\,I(m',n'),
$$
with
$$
H(m,n;m',n')=|h(m,n;m',n')|^2,
$$
so that the processor implements a linear map
$$
O=A'I, \qquad A'_{((m,n),(m',n'))}=H(m,n;m',n')\ge 0.
$$
For arbitrary real, nonnegative intensity transforms, the reported degrees-of-freedom condition is
$$
N\gtrsim 2\,N_i\,N_o,
$$
where $N$ is the total number of optimizable phase-only diffractive features. The work reports that MSE falls rapidly as $N\to 2N_iN_o$ and saturates for $N\gtrsim 2N_iN_o$; shallow networks with $K=1,2$ fail even at $N\gg 2N_iN_o$; deeper stacks with $K\ge 3$ fully unlock the available degrees of freedom; and indirect training yields lower MSE when $N\gtrsim 2N_iN_o$. The paper frames these results as the first demonstration of universal linear intensity transformations under spatially-incoherent illumination.

A learned, optics-driven variant appears in "MetaScope: Optics-Driven Neural Network for Ultra-Micro Metalens Endoscopy" [2508.03596]. There, OIA rectifies intensity decay by learning optical embeddings from two priors: a channel prior $T\in\mathbb{R}^L$ encoding per-wavelength focusing efficiency and a spatial prior $Y_{sp}\in\mathbb{R}^{H\times W}$ encoding radial attenuation. The optical motivation is stated as wavelength-dependent meta-atom efficiency together with radial vignetting, producing an approximate sensor intensity
$$
I_{\mathrm{total}}(\theta,\lambda)=\eta_{\mathrm{meta}}(\theta,\lambda)\cdot \cos^4\theta \cdot I_0(\lambda).
$$
The learned embeddings are
$$
\mathbf{E}_{ch}=\mathcal{F}^{Enc}_{Fc}(T), \qquad
\mathbf{E}_{sp}=\mathcal{F}^{Enc}_{Conv}(Y_{sp}\oplus C_x\oplus C_y),
$$
which modulate channel and spatial attention:
$$
\mathrm{Attn}_{ch}(X)=\sigma\!\left(\mathcal{F}^{Proj}_{Fc}\bigl(Mean_{sp}(X)+\mathbf{E}_{ch}\bigr)\right),
$$
$$
\mathrm{Attn}_{sp}(X)=\sigma\!\left(\mathcal{F}^{Proj}_{Conv}\bigl(Mean_{ch}(X)+\mathbf{E}_{sp}\bigr)\right).
$$
The adjusted feature is formed by sequential channel and spatial gating with a residual skip:
$$
X_{\mathrm{OIA}}=\mathrm{Conv}\Bigl(\mathrm{Attn}_{sp}(X)\odot [\,\mathrm{Conv}(\mathrm{Attn}_{ch}(X)\odot X)\,]\Bigr)+X.
$$
OIA is inserted into the encoder blocks of a NAFNet-based encoder-decoder and is trained end-to-end under the restoration and segmentation objectives rather than a bespoke intensity-consistency term. Reported outcomes include a PSNR improvement of $\approx 0.88$ dB for OIA alone over the Meta-baseline, PSNR 34.85 dB for full OIA+OCC versus 34.13 dB without OIA, and an mDICE increase from 91.09% to 91.37% when adding OIA to the OCC-only model.

## 5. Longitudinal intensity shaping in high-intensity optics

In high-intensity laser systems, OIA is used to prescribe on-axis intensity and, in some formulations, on-axis group velocity over an extended focal region. "Axiparabola: a new tool for high-intensity optics" [2106.06802] formulates the design problem as choosing an aspheric mirror so that
$$
I(z)=I_0(z), \qquad z\in[0,\delta],
$$
and, if required,
$$
v_g(z)=v_{\rm target}(z).
$$
With radial coordinate $r\in[0,R]$ and mirror sag $s(r)$, the paraxial construction is based on the mapping from mirror radius to axial focus coordinate and on the relation
$$
r(z)=R\sqrt{\frac{1}{P_0}\int_0^z \lambda_z(z')\,dz'},
$$
with on-axis irradiance
$$
I(z)=k\,\lambda_z(z)\propto \frac{1}{|dr/dz|}, \qquad k=\frac{2\pi}{\lambda}.
$$
The paraxial sag equation is
$$
\frac{d\sigma}{dr}\approx -\frac{r\,z(r)}{2f_0^2}, \qquad
s(r)=\frac{r^2}{2f_0}+\int_0^r \frac{d\sigma}{dr'}\,dr'.
$$
Two worked examples are given. For a constant-intensity focal line, one obtains
$$
r(z)=R\sqrt{z/\delta}, \qquad f(r)=f_0+\delta\,(r/R)^2,
$$
with example parameters $f_0=400$ mm, $\delta=15$ mm, and $R=38.1$ mm. For the constant-energy focal line relevant to plasma-waveguide generation, the empirical fit used is
$$
f(r)=f_0+0.1\,\delta\,(r/R)+0.9\,\delta\,(r/R)^{1/2}.
$$
Upstream spatio-temporal couplings $\tau(r)$ are then used to tailor the effective on-axis group velocity.

"Programmable Focal Elongation and Shaping of High-Intensity Laser Pulses using Adaptive Optics" [2509.23294] realizes a related programmatic shaping by applying an adaptive phase mask
$$
E_{\rm in}(x,y)=A(x,y)\exp[i\phi(x,y)],
$$
with
$$
A(r)=A_0\exp[-(r/w)^p], \qquad
\phi(r)=\sum_{n=2,4,6,\dots} k\,\alpha_n\,Z_n^0(r/R).
$$
After an off-axis parabola of focal length $f_0$, the total phase is
$$
\Phi(r)=-\frac{k\,r^2}{2f_0}+\phi(r),
$$
and the on-axis intensity follows from the Fresnel integral. Representative simulated lineouts are reported as FWHM$_\ell \approx 2.2$ mm without AO and FWHM$_\ell \approx 35$ mm with aberration $(\alpha_2,\alpha_4,\alpha_6)=(-0.7,-1.2,0.5)\,\mu$m. Experimentally, with $\alpha_2=-0.69\,\mu$m, $\alpha_4=-1.17\,\mu$m, and $\alpha_6=+0.53\,\mu$m, the focal lengthening reached FWHM$_\ell \approx 35$ mm, corresponding to $\simeq 16\times z_R$ for a native Rayleigh length $z_R=2.2$ mm, with on-axis intensity variation below $\pm 10\%$ over the central 30 mm and agreement with simulation within $<5\%$ error in FWHM$_\ell$ and on-axis lineouts. The stated applications include HOFI waveguide generation and flying focus for dephasingless laser-wakefield acceleration.

## 6. Limitations, misconceptions, and outlook

Several limitations recur across OIA formulations. In I-AO, sensorless intensity modes are not strictly orthogonal, superposition can be imperfect, finite SLM phase-to-intensity calibrations limit dynamic range, and additional phase errors introduced by intensity correction must be removed by the DM [2405.16289]. In incoherent diffractive processors, depth matters: shallow networks fail even at large feature counts, which indicates that feature count alone does not guarantee realizability of a target intensity map [2303.13037]. In metalens endoscopy, OIA does not operate as a standalone correction term but is learned jointly with restoration and segmentation, so its effect depends on the surrounding architecture and supervision [2508.03596].

A second misconception is that OIA can replace phase correction. The adaptive-optics literature states the opposite: in both sensor-based and sensorless I-AO, a DM phase routine follows intensity adjustment, and the reported gains are used to show that phase-only AO is insufficient when intensity errors dominate, not that phase is dispensable. Likewise, physics-based inversion methods such as light-in-flight imaging do not directly manipulate light; they recover true intensity by dividing out geometric-optics and scattering effects, so the adjustment is inferential rather than actuated [2103.12464].

The outlook in the literature is correspondingly heterogeneous. Proposed optimizations for I-AO include zonal or data-driven intensity correction bases, integration of Shack-Hartmann wavefront sensing in loop 1 for concurrent $\phi$ and $I$ measurement, and machine-learning accelerated mode selection and metric evaluation. Broader scenarios explicitly identified for OIA include astronomical telescopes with segmented mirrors, free-space optical links with atmospheric scintillation, laser machining through non-uniform media, visual processors operating under natural spatially-incoherent light, and biomedical imaging systems that must compensate structured intensity decay or sample-specific contrast variations [2405.16289] [2408.02703]. This suggests that OIA is best understood not as a single technique but as a general optics-informed design pattern for restoring, reshaping, or learning intensity in optical systems whose dominant errors are not purely phase aberrations.

Source: https://www.emergentmind.com/topics/optics-informed-intensity-adjustment-oia