Charge-Conserving Retinex Decomposition for Time Projection Chamber Images

Particle detectors called time projection chambers record a charged particle's path as an image. One axis is the readout wire, the other is arrival time, and each pixel's brightness measures how much ionization charge arrived. Rare-event searches, such as neutrino detectors or experiments looking for neutrinoless double-beta decay, separate the few true signals from much more common backgrounds by reading fine details of these images. The most important detail is the dense charge "blob" at the end of a stopping track, called a Bragg peak. The images are corrupted in three different ways. Electronics add noise. Charge is lost smoothly as electrons drift farther. The wire's response blurs and reshapes each pulse. Low-light photography has a mature tool for this kind of problem: perturbation-aware Retinex decomposition. It splits an image into an intrinsic part (reflectance), a smooth multiplicative field (illumination), and an explicit noise term. The noise term acts as an escape valve, so corruption is not forced into the parts that matter.

We propose to adapt this decomposition to detector images and to change it in the one way the physics demands. The noise term must be forbidden from carrying net charge. The intrinsic component should then be the deposited energy pattern itself, invariant across detector conditions. Our claim is that this constraint lets a decomposition trained on unpaired real data remove detector-specific corruption while preserving the endpoint charge that background rejection depends on. Unconstrained generative translation cannot guarantee this, and neither can a denoiser trained only on simulation.

Core Claim and Model

A detector image $I$ (wires by time ticks) is modeled as

$$I = h \ast (L \odot R) + N,$$

where $R$ is the deposited ionization charge per pixel (the quantity physicists want), $L$ is a smooth positive gain field (electron-lifetime attenuation $e^{-t/\tau}$, wire-to-wire gain, space-charge yield variations), $h$ is the wire field-response kernel (bipolar on induction planes), $\ast$ is convolution along wires and time, and $N$ is electronic noise. The difference from photographic Retinex is the response convolution $h$, together with a constraint on $N$. Photographic models let noise absorb whatever the reconstruction cannot explain. Here we require that $N$, averaged over any window the size of a Bragg peak, carries approximately zero charge. We also require its power spectrum to match noise measured from signal-free regions of the same data. Detector data contain large signal-free areas, so this noise model comes from real noise rather than an assumed Gaussian. It also avoids the neighbor-pixel downsampling that DEnet uses to build independent noise views, which its authors acknowledge removes fine structure. In a detector image, that structure is exactly the Bragg peak.

What Prior Work Supplies and Leaves Open

The method draws on three low-light papers. Structure-Revealing Low-Light Image Enhancement Via Robust Retinex Model introduced jointly estimating reflectance, illumination and an explicit noise map by regularized optimization. We keep that three-way split but add a known convolutional response and a charge-conservation constraint. Retinexformer supplies a single-stage architecture whose attention is guided by the estimated illumination map, sending restoration capacity to the darkest, noisiest regions. We reuse that guidance so that capacity goes to long-drift, low-gain regions. DEnet supplies the unsupervised principle: two views of one scene under different lighting must yield the same reflectance. We replace its random gamma perturbation, whose bounds the low-light literature reports as fragile, with physically sampled lifetime and gain draws.

On the detector side, the noise study Noise Characterization and Filtering in the MicroBooNE Liquid Argon TPC motivates the work. It shows that static frequency masks and correlated-channel subtraction raise peak signal-to-noise on minimum-ionizing tracks from 6–19 to 16–38. The same study shows that burst noise near 900 kHz and regulator ripple resist linear filtering without clipping small induction pulses. Unpaired image translation to mitigate domain shift in liquid argon time projection chamber detector responses is the closest existing work. It translates between a simplified and a realistic wire response with CycleGAN-family models, and its best model, UVCGAN, cuts downstream electron-count error by up to 80%. That translation is a black-box mapping with no conservation law. The proposal differs by translating through an explicit $R$, which allows charge conservation to be checked and enforced. Vertex-finding and reconstruction of contained two-track neutrino events in the MicroBooNE detector shows that segmentation on raw images fails for short, low-energy tracks. Those tracks are the ones whose endpoint charge is most at risk. The topological analysis in Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment constrains the design: blob charge and end-track twist assume clean hits, so any preprocessing that softens endpoints directly raises false signals.

Research Program

We will build a U-shaped transformer with three output heads. One predicts $R$ at full resolution with no downsampling in the final stage. One predicts a low-resolution $L$ that is upsampled smoothly. One predicts $N$. The known or parametrized kernel $h$ re-renders the image from these outputs. Training combines four losses.

The first is reconstruction of $I$. The second is DEnet-style invariance: the same simulated deposit, rendered under two sampled $(L, h)$ settings, must produce the same $R$. The third is a spectral match between predicted $N$ and real signal-free noise patches. The fourth is the charge-conservation penalty, which keeps the local window sums of $N$ near zero.

Translation between detector domains then becomes a simple procedure: decompose, swap $h$ and $L$, re-render. On real, unlabeled data the invariance and noise losses still apply, so the model can adapt to the real detector without truth labels. The analysis question is whether each structural piece earns its place, especially the charge constraint. We will answer it with ablations, not assume it.

Baselines

The naive reference is a downstream estimator trained on the source domain and applied unchanged to target-domain images. This is the error UVCGAN's 80% reduction was measured against, and we will reproduce it on the same downstream estimator.

The classical reference is detector signal processing in the style of the MicroBooNE noise paper: correlated-noise subtraction, frequency masking, and one-dimensional deconvolution of the field response. It is the strongest competitor on charge fidelity because it is linear and roughly charge-preserving by construction. We will run it with its filters tuned on the same signal-free noise patches our model sees.

The strongest existing learned method is UVCGAN. It reports up to 80% reduction in downstream electron-count error for realistic-to-simplified translation on SLATS, according to Unpaired image translation to mitigate domain shift in liquid argon time projection chamber detector responses. We will retrain it with its published configuration on identical splits, matching GPU-hours and hyperparameter search budget. Beating it means more than exceeding 80% on aggregate electron count. It also means lower error on charge in the last few centimeters of stopping tracks, where we expect generative smoothing to concentrate.

A fourth baseline is a supervised U-Net denoiser trained on simulated deposits with real noise patches overlaid, given the same compute. This is the most dangerous competitor (see below).

Ablations remove, one at a time, the charge-conservation constraint, the separate $L$ field, the invariance loss, and the measured-noise spectral loss. If removing the charge constraint does not change endpoint charge error, the central claim fails.

Why not just…?

Why not just use UVCGAN, which already works on these images? Adversarial and cycle losses reward images that look like the target domain. They do not reward correct charge, so a softened or invented blob is not penalized if it looks plausible. The aggregate electron-count metric could hide endpoint errors. We will report endpoint-window charge error and the false-blob rate on single-electron tracks separately. If UVCGAN matches us there, our advantage reduces to interpretability, and we will say so.

Why not just keep classical filtering and deconvolution? It conserves charge well and is cheap. However, the MicroBooNE study documents that it leaves burst and regulator noise in place, and that it clips low-amplitude induction pulses. The decisive comparison is on images with injected real burst-noise segments, where we measure induction-plane charge error and short-track detection efficiency. If classical processing ties us, the learned model is not justified.

Why not just train a supervised denoiser on simulation plus overlaid real noise? This is a real threat. When the only gap between simulation and data is noise, overlay training is essentially the right answer, and our model should not beat it. Our advantage can appear only when the response $h$ or gain $L$ also differs between simulation and data, which is SLATS's realistic-versus-simplified setting. We therefore run this baseline on two regimes: noise-only mismatch and response mismatch. We predict parity in the first and a gap in the second. If there is parity in both, the decomposition adds nothing beyond noise modeling.

Proposed Experiments

The primary benchmark is SLATS. Its test set pairs the same tracks in both response domains, so pixel-level and charge-level truth exist. If SLATS images lack realistic electronic noise, we overlay noise with spectra drawn from the published MicroBooNE noise characterization's categories, using the same injection for all methods. Metrics are pixel error against the paired target, total and endpoint-window charge error, and the downstream electron-count error used by the UVCGAN study.

The secondary test asks whether a background rejection variable survives. On simulated stopping tracks we compute a Bragg-peak blob-charge discriminator after each method. We then compare its separation of two-blob from one-blob topologies, analogous to PandaX-III's $0\nu\beta\beta$ versus gamma cut. The figure of merit is background acceptance at fixed signal efficiency. Simulated statistics cannot reach the $10^7$ suppression PandaX-III reports, so we compare relative rejection at accessible working points.

The decisive check is the charge-constraint ablation on endpoint charge and false-blob rate. If it moves those numbers and the overlay baseline loses only under response mismatch, the proposal's claim holds. Otherwise it does not.