When a dark image is brightened, every grain of sensor noise is brightened with it. Some of that grain then looks like structure: false edges, false corners, false texture. Downstream systems such as trackers, visual odometry and pose estimators latch onto these "phantom features" and fail. Particle physicists face the same problem when searching for rare events in a noisy detector. Their standard answer is delayed coincidence detection. They accept an event only if a second, physically linked signal appears inside a tight time-and-space window. They then measure how often pure noise passes that test by repeating the search in a deliberately mismatched "off-window." That measured accidental rate is subtracted, or used to set the cuts, so the false-positive rate is known rather than guessed.
This note proposes carrying that whole procedure, including the accidental-rate measurement, into low-light enhancement. Each short exposure is split into two sub-exposures whose noise is independent. A brightened detail or keypoint is trusted only where it appears in coincidence in both. The rate of chance coincidences is measured directly by pairing the sub-exposures with a small deliberate spatial offset. The enhancer's local gain and the feature detector's threshold are then set so that the measured fraction of chance coincidences stays below a chosen target. The result is an enhancer whose phantom-feature rate is a dial with a measured value, not a side effect.
Core Claim
Current low-light enhancers decide how much detail to amplify from a learned or hand-set prior on a single image. Retinex methods model the image as reflectance times illumination plus noise. They recover reflectance by dividing by illumination, which scales noise by $1/L$ in the darkest regions, and they then fight the blow-up with spatial smoothing. This happens in Deep Retinex Decomposition for Low-Light Enhancement (BM3D on reflectance), Joint Enhancement and Denoising Method via Sequential Decomposition (total-variation illumination, then weighted reflectance), URetinex-Net (unfolded optimization with learned priors) and CERL (alternating enhancement and self-supervised denoising). These papers motivate the proposal: they document the visibility-versus-fidelity tradeoff. None of them can report how many of the details it amplified are real.
Our claim is that a two-view coincidence test with off-window accidental subtraction gives that number cheaply. It can also be used to choose gain per region. The expected effect is that phantom keypoints drop sharply at matched true-keypoint recall, compared with denoise-then-detect pipelines. The gate should also come with a calibrated false-feature rate that transfers across sensors without retuning.
The statistics follow the physics directly. Suppose that in a local region a detector, such as an edge or corner response above a threshold, fires falsely on noise with per-pixel probability $p$ in each sub-exposure. Suppose it fires on true structure with probability $q$. Coincidence within a radius $r$ then passes noise at a rate of roughly
$$A \approx p^2 \pi r^2,$$
while true structure passes at roughly $q^2$. This is the imaging analogue of the accidental rate $R^2 \Delta t$ used for inverse beta decay in Daya Bay and KamLAND. The squared false rate against the linear window size is what makes coincidence powerful. The off-window estimate $\hat A$ comes from pairing sub-exposure one at location $x$ with sub-exposure two at $x+d$, where $|d| \gg r$ but $d$ is small enough that local noise and texture statistics are unchanged. In a region with $C$ observed coincidences, the purity is $(C-\hat A)/C$. The gate picks the largest gain and lowest threshold for which this purity stays above $1-\alpha$.
Research Program
The first component is the split. On raw sensors, a burst of short frames gives sub-exposures with independent shot and read noise; we pair alternate frames. For pure photon-count data, binomial thinning of a Poisson count gives two exactly independent Poisson halves. This lets us build controlled experiments from single raw captures. We will use it only for analysis, because real read noise breaks exact independence. Rendering always uses the summed exposure, so splitting costs no photons in the final image. The split is used only to decide what to trust.
The second component is the gate itself. We compute a structure map in each sub-exposure using the enhancer's own high-frequency residual and a standard keypoint detector. We form the coincidence map within radius $r$ and estimate accidentals from several random offset directions, taking the median. The median guards against periodic textures such as stripes, where a shifted copy of the signal can coincide genuinely. The per-region purity then controls a gain map. It also drives a blend that keeps enhanced high-frequency detail where coincidences are pure and falls back to a smooth illumination-only brightening elsewhere. The gate is a wrapper: it can sit around any enhancer, and we will wrap both a curve-based one and a Retinex-based one.
The third component deals with correlated backgrounds. Fixed-pattern noise and hot pixels appear identically in both sub-exposures, so they produce real coincidences that are not scene structure. This is the analogue of cosmogenic fast neutrons, which fake both stages of inverse beta decay. We follow the physicists' answer, which is to measure correlated backgrounds in source-off runs. Here that means subtracting a coincidence map computed on dark frames captured with the same settings. Without this step the method will fail on real sensors, and the ablations will show by how much.
The proposal deliberately does not build a heavy spatio-temporal transformer. Separating the Wheat from the Chaff already shows that attention over photon-arrival histograms can find time-of-flight coincidence peaks in active single-photon depth sensing (RMSE 0.0711 on Middlebury at signal-to-background 2:50). That work supplies evidence that coincidence structure is learnable. It targets depth from active illumination, not passive intensity enhancement, and it gives no calibrated false-positive rate. Time-gated fluorescence imaging, reviewed in Enhancing biological imaging with TADF probes, rejects short-lived autofluorescence by hardware gating. It motivates the "wait for a second, independent confirmation" logic, but it needs engineered emitters, which passive cameras lack.
Baselines
The naive reference is plain brightening of the summed exposure followed by the downstream detector or tracker, with no denoising. This sets the floor for phantom features. Zero-DCE serves as the realistic version of this floor: it is fast, has no explicit denoising step, and is reported in CERL's comparison to leave heavy grain in dark regions.
The strongest competitor on raw bursts is the permutation-invariant burst denoiser of Burst Denoising of Dark Images. It fuses up to ten short raw frames by set pooling and reports PSNR 29.93 and SSIM 0.899 (LPIPS 0.429) on the SID Sony test set with ten-frame bursts. It is the strongest because it already exploits multiple independent noise draws, which is exactly the resource our gate uses. Any gain we show must come from the coincidence-and-calibration logic, not from extra frames. We will run it on identical bursts, at identical burst length, retrained on the same SID split with the same compute budget. Its output will then go to the same feature detector. Beating it means lower phantom-keypoint rate at equal true-keypoint recall, with PSNR no more than a few tenths of a decibel worse. We do not expect to beat it on PSNR, and we will not claim to.
For single-frame joint enhancement and denoising, the reference is CERL, which reports PSNR 20.05 on its real mobile benchmark (RLMP) against 18.68 for DRBN. We will wrap CERL in our gate, so the comparison isolates the gate.
For downstream tasks the references are task-specific. On UAVDark135, Tracker Meets Night reports a 13.3% success and 15.4% precision improvement with its enhancer. On DARK FACE, R2RNet reports that its enhancement raises DSFD face detection AP from 17.12% to 33.98%. On ExLPose-OCN, the dual-teacher method of Domain-Adaptive 2D Human Pose Estimation reports 37.6 AP.
Two ablations isolate the key ingredient. The first keeps the coincidence test but replaces off-window calibration with a single fixed threshold tuned on validation data. This asks whether measuring the accidental rate matters, or whether coincidence alone does the work. The second keeps calibration but drops the split. It sets thresholds from a fitted single-frame noise model, asking whether independent views matter or whether a good noise model suffices. A third ablation removes dark-frame subtraction to size the correlated-background problem.
Why not just…?
Why not just run a strong burst denoiser and then detect features? This is the most serious threat, because a burst denoiser sees the same independent noise draws. Its training objective is pixel fidelity to a long-exposure reference. Phantom structures that resemble real texture are cheap under that loss, and the denoiser provides no estimate of how many survive. The SID comparison measures phantoms directly against features detected on the long-exposure reference. If the Karadeniz denoiser plus a standard detector already matches our phantom rate at equal recall, the remaining contribution shrinks to calibration. We would then report the gate as a confidence-estimation layer on top of burst denoising, not as a better enhancer.
Why not just require features to persist across consecutive video frames, as trackers already do? Persistence checks are coincidence tests too, but across frames that contain motion. The search radius must then grow with motion, and by the $p^2 \pi r^2$ relation the accidental rate grows with the square of that radius. This is the same noise-and-motion dilemma that Coherent Event Guided Low-Light Video Enhancement and Dynamic Low-Light Imaging with Quanta Image Sensors document for alignment. Splitting one exposure keeps $r$ near one pixel. We will test this directly on SID bursts with synthetic camera motion of increasing magnitude. Our prediction is that frame-persistence gating matches ours at zero motion and degrades as motion grows. If it does not degrade, the split is unnecessary.
Why not just set detector thresholds from a calibrated noise model of a single frame? This is the second ablation. Real low-light noise includes signal-dependent shot noise, row noise and fixed pattern noise that depart from simple models, as both CERL and the sensor-noise literature stress. A model-based threshold therefore inherits model error, whereas off-window accidentals are measured on the actual capture. If a well-fitted noise model gives the same calibration error on held-out sensors, our measurement step is redundant and we will say so.
Proposed Experiments
The primary testbed is SID Sony raw data from Learning to See in the Dark, with short-exposure bursts as sub-exposures. Keypoints and edges detected on the long-exposure reference serve as ground-truth structure. We report phantom-keypoint rate, true-keypoint recall, and repeatability of short-exposure keypoints against the reference, alongside PSNR, SSIM and LPIPS. The decisive check is calibration. Across a sweep of target purities $1-\alpha$, the gate's measured purity $(C-\hat A)/C$ should match the purity measured against the reference. We plot one against the other. A method whose predicted and actual phantom rates agree on the Sony camera, and still agree on the Fujifilm subset without retuning, has delivered what the physics analogy promises. A poor match means the off-window estimate is biased, most likely by texture or fixed-pattern noise, and the dark-frame ablation will show which.
Downstream, we wrap the same tracker used by Tracker Meets Night on UAVDark135 and DarkTrack2021, using consecutive frames as an approximate split. We report success, precision and frames per second on the same embedded hardware class. We wrap DSFD on DARK FACE and compare against the R2RNet enhancement result. We wrap a pose network on ExLPose-OCN, where the dual-teacher paper notes that enhancement artifacts can hurt more than darkness itself. That makes ExLPose-OCN the setting where removing phantoms should matter most.
For thermal visual odometry, Passive Infrared Image Detail Enhancement with Points and Lines is the motivating case for phantom-driven drift. The file names no public dataset for it, so we treat it as a qualitative replication target. On LOL and LSRW, which are processed 8-bit images without bursts, the method cannot run as designed. We will use them only for a no-regression check, applying the gate to the single image with a learned split and reporting whether the PSNR reported by URetinex-Net (21.32) and R2RNet (20.207) is preserved.
References
- Deep Retinex Decomposition for Low-Light Enhancement
- Joint Enhancement and Denoising Method via Sequential Decomposition
- URetinex-Net: Retinex-based Deep Unfolding Network for Low-light Image Enhancement
- CERL: A Unified Optimization Framework for Light Enhancement with Realistic Noise
- Measurement of the Reactor Antineutrino Flux and Spectrum at Daya Bay
- High Sensitivity Search for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mover accent="true"><mml:mi>ν</mml:mi><mml:mo>¯</mml:mo></mml:mover><mml:mi>e</mml:mi></mml:msub></mml:math>’s from the Sun and Other Sources at KamLAND
- Separating the Wheat from the Chaff: Spatio-Temporal Transformer with View-interweaved Attention for Photon-Efficient Depth Sensing
- Enhancing biological imaging with high-performance probes based on thermally activated delayed fluorescence
- Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement
- Burst Denoising of Dark Images
- Tracker Meets Night: A Transformer Enhancer for UAV Tracking
- R2RNet: Low-light image enhancement via Real-low to Real-normal Network
- Domain-Adaptive 2D Human Pose Estimation via Dual Teachers in Extremely Low-Light Conditions
- Coherent Event Guided Low-Light Video Enhancement
- Dynamic Low-Light Imaging with Quanta Image Sensors
- Learning to See in the Dark
- Passive Infrared Image Detail Enhancement with Points and Lines for Visual Odometry