Stopping a Single-Pixel Camera as Soon as the Photons Prove a Change

A single-pixel camera forms an image by shining a sequence of patterns onto a scene, for example from an array of tiny tilting mirrors, and recording one total light reading per pattern with a single sensitive detector. Because the patterns arrive one at a time, the time to capture an image grows with the number of patterns. In very dark scenes, where the detector counts individual photons, each reading is also noisy in the way coin-flip counts are: its relative uncertainty falls only as $1/\sqrt{N}$ in the number of photons $N$. Current systems accept both costs. They project a fixed, large number of patterns, such as 1,800 patterns over about 53 seconds in Single photon counting imaging system via compressive sensing, then reconstruct an image, and only after that ask whether anything interesting is in it.

We propose to run the statistical decision while the patterns are still being projected and to stop the camera the moment the photons already counted are enough to reject "nothing has changed" at a guaranteed false-alarm rate. The tool is an e-process: a running product of likelihood ratios that, under the null hypothesis, behaves like the wealth of a gambler in a fair game. Its false-alarm guarantee therefore holds at whatever moment we choose to stop. Because time order already separates past photons, which we use to guess what changed, from future photons, which we use to test that guess, no photon has to be held back for validation. The same running guess also picks the next pattern, concentrating light where the change appears to be. The claim is that, for detection-type tasks, latency is set by how quickly evidence accumulates rather than by how many patterns full reconstruction would need.

Core Claim

Take a known reference scene $x_{\text{ref}}$, such as an empty corridor, a defect-free part, or a calibrated background. Let $a_t$ be the non-negative pattern projected at step $t$, $\tau$ the dwell time per pattern, and $d$ the dark-count rate of the detector. The null hypothesis is that the photon count $y_t$ is Poisson with mean $\lambda^0_t = \tau(a_t^\top x_{\text{ref}} + d)$. Let $\hat{x}_{t-1}$ be any estimate of the scene built only from counts before step $t$; it can come from a neural network, a Poisson maximum-likelihood solver, or anything else. The test statistic is

$$E_t = \prod_{s=1}^{t} \frac{\text{Pois}\big(y_s;\ \tau(a_s^\top \hat{x}_{s-1} + d)\big)}{\text{Pois}\big(y_s;\ \lambda^0_s\big)}.$$

Under the null, $E_t$ is a non-negative martingale with mean one, which here simply means that each new factor has expected value one given the past. This holds whatever $\hat{x}$ is and however $a_t$ was chosen from the past. By Ville's inequality, the chance that $E_t$ ever exceeds $1/\alpha$ is at most $\alpha$. We stop and declare a change when $E_t \ge 1/\alpha$, and we declare "no change detected" when the photon or time budget runs out.

Wald's classical argument gives the expected number of patterns needed to detect a true change: roughly $\log(1/\alpha)$ divided by the average per-pattern Kullback–Leibler divergence (a measure of how distinguishable two count distributions are) between the alternative and null count distributions. That is the latency lever. We therefore pick the next pattern to maximize this predicted divergence,

$$a_t = \arg\max_{a \in \mathcal{A}} \Big[\lambda^1 \log\frac{\lambda^1}{\lambda^0} - \lambda^1 + \lambda^0\Big],\quad \lambda^1 = \tau(a^\top \hat{x}_{t-1} + d),\ \lambda^0 = \tau(a^\top x_{\text{ref}} + d),$$

where $\mathcal{A}$ is the set of binary patterns the mirror device can display, restricted to a tractable family such as blocks and Hadamard rows masked to candidate regions.

This choice directly addresses the two Poisson penalties identified in Compressed sensing performance bounds under Poisson noise. That paper shows that, for a fixed photon budget, spreading light over more measurements raises the error bound (the "dilution effect"). It also shows that the positive offset needed to make patterns non-negative adds background counts and therefore variance. A pattern chosen to maximize divergence spends photons only on a few highly discriminating measurements. It also tends to be sparse, switching off mirrors that view regions where the guess matches the reference, which shrinks the background-count penalty. What would be new is the combination: decision-driven stopping with false-alarm control that holds under arbitrary stopping, applied to the raw photon stream, with patterns chosen from the same running statistic.

Relation to Prior Work

Two lines of work motivate the gap. Latency-reduction methods attack the problem with hardware parallelism or faster decoders, all at a fixed measurement count. FPA-CS splits the mirror array across a 64×64 sensor to reach 32 fps, LiSens uses a line sensor to reach a 1 MHz measurement rate, and Deep learning for real-time single-pixel video learns patterns and a fast decoder to reach 30 fps from 333 patterns per frame. None of them decides when to stop. All of them spread light more thinly, which is harmful when photons are scarce. The adaptive methods change which patterns are used rather than how many. Reconstruction-free Cascaded Adaptive Compressive Sensing allocates samples from raw measurements but was built for high-signal conditions. Pre-Capture Privacy via Adaptive Single-Pixel Imaging closes a pattern-generation feedback loop for privacy, with no noise model or stopping rule. We reuse their feedback-loop idea but drive it with a statistical objective.

Two papers supply the method, and one of them constrains it. Hypothesis Testing in Imaging Inverse Problems controls Type I error on reconstructed images by splitting one Gaussian observation into a part that forms the hypothesis and a part that tests it. The authors note that this split costs signal-to-noise ratio, which is exactly the resource that is scarce in our setting. Our design replaces the split with time ordering, so every photon serves first as test evidence and then as input to future guesses. Approximation Algorithms for Active Sequential Hypothesis Testing gives greedy information-driven test selection over abstract action spaces with logarithmic test counts. We instantiate that greedy principle with Poisson divergence over physically realizable mirror patterns. Photon-Starved Scene Inference using Single Photon Cameras constrains the design: off-the-shelf "denoise then decide" pipelines collapse at 0.11 photons per pixel. We therefore never require a clean reconstruction. The guess $\hat{x}$ only needs to point in roughly the right direction, because the test's validity does not depend on its quality.

Research Program

The first component is the test itself, implemented in simulation for a single-pixel camera with photon counting. The simulator models non-negative binary patterns, dwell time, dark counts, and the 50% light loss from mirrors in the off state. For the running guess $\hat{x}$ we compare three predictors. The first is a few warm-started iterations of Poisson maximum likelihood with total-variation regularization, in the style of SPIRAL. The second is a small feed-forward network trained with a Poisson rather than squared-error loss on photon-scale-space augmentations borrowed from Photon Net. The third is a deliberately poor predictor, used to confirm that validity survives a bad guess.

The second component is pattern selection, which maximizes the divergence objective above over a structured pattern family. Its compute is budgeted so that selection runs within one mirror cycle; otherwise the latency gain is fictitious.

The third component extends the method beyond a known reference. One extension handles small reference uncertainty, such as illumination drift, by using a composite null that is a small family of reference scenes, taking the worst-case denominator. The other is a K-class decision, where per-class generative models replace $x_{\text{ref}}$. In that case validity becomes conditional on the class models, and we will report it as such rather than claiming an exact guarantee.

Baselines

The naive reference point is fixed-length capture. The full schedule is projected, for example the 333 learned patterns of Deep learning for real-time single-pixel video or a 1,800-pattern random schedule as in Single photon counting imaging system via compressive sensing. The image is then reconstructed with Poisson maximum likelihood, and a likelihood-ratio test against $x_{\text{ref}}$ is applied, thresholded to give exactly $\alpha$ false alarms at that fixed length. This baseline is valid and simple, and it shows what latency costs without adaptivity. To beat it, our method must reach the same detection power at matched $\alpha$ and matched total photon budget using substantially fewer patterns. Our target is a median reduction of at least 3×, set before running.

The strongest existing statistical competitor is the measurement-splitting test of Hypothesis Testing in Imaging Inverse Problems. It reports 73.38% power at 5% significance on Oxford 102 Flowers versus 58.22% for a zero-shot CLIP baseline, under Gaussian noise. We found no reported result on Lacuna for this method under Poisson counting noise. We port it fairly by binomial thinning: randomly assigning each counted photon to split one or split two gives two independent Poisson observations, so the method's validity carries over exactly. We run it on the same simulated scenes, photon budgets, predictor networks, and hyperparameter search budget as our method. Beating it means higher power at equal $\alpha$ and equal photons, or equal power with fewer photons.

The ablation that isolates the key ingredient is our e-process run with a fixed, non-adaptive Hadamard or random pattern order. This is essentially a sequential probability ratio test with a learned plug-in alternative. It separates the gain from anytime stopping from the gain from divergence-driven pattern choice. A second ablation replaces the Poisson count model with a Gaussian approximation in both test and selection, to measure the cost of ignoring count statistics at low flux.

Why not just…?

Why not just expose longer per pattern and accept a lower frame rate? Longer exposure lowers relative noise only as $1/\sqrt{N}$, and it lengthens the window during which motion or occluders corrupt the fixed linear model. Dynamic occlusion removal in single-pixel imaging system based on self-check reports reconstruction error rising from 0.0028 to 0.0246 when an occluder enters during capture. Our fixed-length baseline, run at several dwell times, quantifies this directly. We expect early stopping to dominate it whenever changes are localized, because evidence accumulates at the divergence rate rather than at the reconstruction rate.

Why not just threshold a classifier's confidence as patterns arrive and stop when it is high? This is the cheapest early-stopping scheme, and it will often stop early. However, repeatedly checking a fixed-sample threshold inflates false alarms, and network confidence is poorly calibrated at sub-photon-per-pixel levels, where Photon Net reports 42.37% versus 27.21% accuracy at 0.11 photons per pixel on CUB-200-2011. We include this baseline and report its empirical false-alarm rate under the null across stopping times. If it stays below $\alpha$ in practice, our contribution narrows to the guarantee, and we will say so.

Why not just run a classical sequential probability ratio test with standard patterns? This is the real threat. For a simple alternative, Wald's test is already optimal, and a plug-in version may capture most of the latency gain without any pattern adaptation. The fixed-order ablation answers this question exactly. If adaptive patterns do not reduce expected stopping time at equal $\alpha$ by a clear margin, especially for small or low-contrast changes where dilution should matter most, then adaptivity is not worth its compute. The proposal then reduces to an anytime-valid sequential test on single-pixel streams, which is still the part no cited system provides.

Why not just use a single-photon camera array and avoid sequential capture? In visible light this is a fair answer, which is why our hardware experiments target settings where arrays are expensive or unavailable, following the cost arguments of FPA-CS. The test also applies unchanged to parallel architectures, where each sub-detector runs its own e-process and the results are combined by product. In those architectures dilution across thousands of pixels makes per-pixel evidence weaker still.

Proposed Experiments

The simulated study uses STL-10 at 128×128, matching the single-pixel video setup above, and Oxford 102 Flowers, matching the hypothesis-testing benchmark. Reference scenes are clean images. Alternatives insert objects, patches, or texture edits of controlled area and contrast. Total photon budgets range from 0.1 to 100 photons per pixel, with realistic dark counts. We report the empirical false-alarm rate under the null at $\alpha \in \{0.05, 0.01\}$, detection power, and the distributions of patterns and photons consumed until stopping. Each is plotted against change area and contrast, the regimes where dilution should separate adaptive from fixed patterns. For the K-class extension we report accuracy versus latency on STL-10 at the same photon levels, compared against fixed-length classification.

The hardware study uses a mirror-array camera with a Geiger-mode photon-counting detector, like the ISSPCCS setup, imaging a static reference target before and after placing small objects. We measure wall-clock latency including pattern-selection compute, and we compare it against the full 1,800-pattern capture.

Three checks decide the outcome. First, the null false-alarm rate must stay at or below $\alpha$ across thousands of null trials for every predictor, including the deliberately bad one. If it does not, the implementation or the null model is wrong. On hardware, this check is where dark-count drift and reference misregistration will show up, and we will calibrate for them and report residual violations. Second, at equal photons and $\alpha$, the e-process must beat thinning-based measurement splitting in power. Third, adaptive patterns must reduce expected stopping time relative to the fixed-order sequential test. Failing the third check would mean the latency gain comes from stopping, not from sensing design, which is a smaller but still publishable result.