PoissonNet: Poisson-Guided Low-Light Enhancement

Updated 4 July 2026

The paper demonstrates that incorporating a Poisson noise model into a Retinex-style decomposition significantly improves enhancement under extreme low-light conditions.
PoissonNet employs a unified encoder–decoder CNN with shared illumination and channel-specific reflectance and noise components, ensuring balanced color constancy and effective denoising.
The model achieves superior PSNR, SSIM, and MAE metrics on the LOL dataset, validating its efficiency in handling signal-dependent shot noise.

PoissonNet is a light-weight deep learning-based Poisson-guided decomposition network for extreme low-light image enhancement and denoising. In the formulation introduced in "A Poisson-Guided Decomposition Network for Extreme Low-Light Image Enhancement" (Rao et al., 4 Jun 2025), it unifies Retinex-based decomposition and Poisson-aware denoising within a single encoder–decoder CNN. The model takes a low-light RGB image $X$ , decomposes it into illumination $L$ , reflectance $R$ , and noise $N$ , reconstructs the observation as $X^c(i) = L(i) \circ R^c(i) \circ N^c(i)$ , and produces the enhanced image $\hat{Y} = L \circ R$ by removing the noise component (Rao et al., 4 Jun 2025). Its central premise is that extreme low-light degradation is dominated by signal-dependent shot noise rather than additive Gaussian noise, so enhancement and denoising should be learned jointly under a Poisson model.

1. Problem setting and motivation

PoissonNet targets extreme low-light images acquired under very poor illumination, low exposure, and small photon counts. In this regime, the signal-to-noise ratio is very low, visibility is poor, contrast collapses, and the dominant corruption is shot noise with Poisson statistics rather than signal-independent Gaussian noise (Rao et al., 4 Jun 2025). Because Poisson noise is signal-dependent, darker regions exhibit larger relative uncertainty, and brightening them without an explicit noise model tends to amplify artifacts.

This motivates a departure from conventional low-light enhancement pipelines. Retinex-based methods such as RetinexNet, LIME, and SRIE model an image as the product of reflectance and illumination, but they generally assume that noise is small or approximately Gaussian. Under extreme low-light conditions, such assumptions lead to failure modes that the paper identifies explicitly: amplified noise after illumination boosting, color distortion caused by per-channel corruption, and unstable separation of reflectance from illumination (Rao et al., 4 Jun 2025). A plausible interpretation is that low-light enhancement becomes ill-posed unless the decomposition itself is conditioned on the underlying photon-counting statistics.

In that respect, PoissonNet belongs to a broader line of Poisson-aware imaging models. Earlier work such as "Deep Convolutional Denoising of Low-Light Images" (Remez et al., 2017) showed that CNNs trained directly on Poisson-corrupted data can outperform variance-stabilization pipelines for photon-limited denoising, but PoissonNet differs by embedding the Poisson prior inside a Retinex-style decomposition rather than treating denoising as a standalone image restoration task.

2. Network architecture and decomposition model

The network is a unified encoder–decoder CNN with three output heads. Its encoder begins with an initial $3 \to 64$ convolution using a $3 \times 3$ kernel, followed by three stride-2 convolution layers for downsampling and feature extraction, with ReLU after each convolution. The decoder uses three transposed-convolution layers and U-Net-style skip connections that concatenate encoder features into the corresponding decoder stages, preserving fine spatial detail while retaining higher-level context (Rao et al., 4 Jun 2025).

At the output, the decoder branches into three parallel heads. The illumination head predicts $L$ with sigmoid activation; the reflectance head predicts the three-channel reflectance map $R$ with sigmoid activation; and the noise head predicts the three-channel noise map $L$ 0 with tanh activation (Rao et al., 4 Jun 2025). The decomposition is written channelwise as

$L$ 1

and the enhanced image is

$L$ 2

A distinctive design choice is that $L$ 3 is shared across color channels, whereas $L$ 4 and $L$ 5 are channel-specific. This shared illumination field is intended to preserve color constancy under ambient illumination, while channelwise reflectance and noise absorb chromatic structure and per-channel corruption (Rao et al., 4 Jun 2025). The paper emphasizes that no ground-truth illumination, reflectance, or noise maps are provided; the model learns the latent decomposition directly from paired low-light and normal-light images.

The decomposition extends the classical Retinex model $L$ 6 by adding a noise factor. Although Poisson noise is not strictly multiplicative, the model treats it as a signal-dependent component in the factorization. This suggests that PoissonNet should be understood less as a literal physical factorization than as a constrained latent-variable parameterization that aligns Retinex decomposition with photon-limited imaging.

3. Objective function and Poisson-guided optimization

Training is driven by a composite loss that couples input reconstruction, target reconstruction, edge-aware illumination regularization, and an explicit Poisson-guided noise prior (Rao et al., 4 Jun 2025). The input reconstruction term is

$L$ 7

which anchors the three-way decomposition to the observed low-light image. The clean-image reconstruction term is

$L$ 8

where $L$ 9 is the paired normal-light target. This forces the illumination–reflectance product to approximate the clean enhanced image.

To regularize the decomposition, the model uses the smoothness preserving smoothness loss

$R$ 0

with $R$ 1 in the experiments (Rao et al., 4 Jun 2025). Here $R$ 2 penalizes illumination variation, while the factor $R$ 3 relaxes that penalty near strong reflectance edges. Operationally, this is an edge-aware smoothness prior for illumination: flat regions are smoothed strongly, whereas structure aligned with reflectance is preserved.

The Retinex estimation loss combines these terms: $R$ 4 with $R$ 5 chosen empirically. The Poisson-specific component is the noise loss

$R$ 6

where $R$ 7 provides numerical stability (Rao et al., 4 Jun 2025). This term teaches the predicted noise map $R$ 8 to approximate a signal-normalized Poisson noise pattern derived from the clean target, thereby encoding the fact that darker regions should exhibit stronger relative noise.

The total training objective is

$R$ 9

The paper states that $N$ 0, $N$ 1, and $N$ 2 are chosen empirically to balance reconstruction, decomposition, and denoising (Rao et al., 4 Jun 2025). Unlike adversarial or staged training schemes, PoissonNet is trained end-to-end with this single composite loss.

4. Training protocol and empirical performance

PoissonNet is trained on the LOL dataset, which contains 500 paired low-light and normal-light images captured by varying exposure time and ISO while keeping other camera settings fixed (Rao et al., 4 Jun 2025). Training uses RGB patches of size $N$ 3 and runs for 100 epochs. The normal-light image $N$ 4 serves both as the supervision target for $N$ 5 and as the basis for Poisson sampling in $N$ 6.

The evaluation compares PoissonNet with LIME, Deep Retinex / RetinexNet, SRIE, Zero-DCE, and a light-weight low-light enhancement model by Ko et al. Four metrics are reported: PSNR, SSIM, MAE, and NIQE (Rao et al., 4 Jun 2025). The paper reports that PoissonNet achieves the best or second-best scores across all metrics, with the strongest gains in PSNR, SSIM, and MAE.

Method	PSNR / SSIM	NIQE / MAE
LIME	15.66 / 0.42	9.35 / 35.57
Deep Retinex	16.06 / 0.58	4.31 / 35.78
SRIE	9.85 / 0.40	7.72 / 80.00
Zero-DCE	12.35 / 0.55	3.42 / 61.71
Light-weight [Ko]	15.99 / 0.59	4.54 / —
PoissonNet	17.12 / 0.67	3.54 / 32.33

The reported outcome is that PoissonNet attains the best PSNR, best SSIM, best MAE, and second-best NIQE, close to Zero-DCE’s NIQE value (Rao et al., 4 Jun 2025). This pattern is consistent with the design goal: the method prioritizes physically plausible denoising and clean reconstruction under severe low-light degradation, rather than only no-reference perceptual scores.

5. Visual behavior, color constancy, and efficiency

The qualitative results emphasize three properties: natural brightness and contrast, effective noise suppression, and minimal color distortion, particularly at the lowest illumination levels (Rao et al., 4 Jun 2025). The paper notes that baseline methods increasingly exhibit yellowish or bluish casts, blotchy artifacts, and unstable enhancement as illumination decreases, whereas PoissonNet remains comparatively stable.

This behavior is tied directly to the decomposition. A single shared illumination map $N$ 7 enforces consistent brightening across RGB channels, while separate channelwise reflectance $N$ 8 and noise $N$ 9 allow chromatic detail and sensor-level corruption to be modeled separately. The smoothness preserving smoothness loss further limits haloing and gradient reversal by coupling illumination regularization to reflectance structure. Skip connections in the encoder–decoder also help maintain texture and edge fidelity (Rao et al., 4 Jun 2025).

The architecture is deliberately light-weight. The network uses only three downsampling stages, three upsampling stages, standard $X^c(i) = L(i) \circ R^c(i) \circ N^c(i)$ 0 convolutions, and no transformers, attention modules, or adversarial discriminator (Rao et al., 4 Jun 2025). The paper therefore presents the method as relatively fast and potentially suitable for resource-constrained deployment. It explicitly identifies mobile cameras, autonomous driving, and surveillance as application settings where the combination of low-light enhancement, denoising performance, and architectural simplicity is practically relevant (Rao et al., 4 Jun 2025).

Within the low-light imaging literature, the paper itself names the method a “Poisson-guided decomposition network” or “Poisson-guided image decomposition framework”; “PoissonNet” is a concise shorthand for that model (Rao et al., 4 Jun 2025). The shorthand is useful, but it is not unique across arXiv. In other subfields, “PoissonNet” or closely related labels refer to a CNN for solving the Poisson equation on Cartesian meshes (Özbay et al., 2019) or to a local-global surface-learning architecture that propagates features by solving Poisson systems on triangle meshes (Maesumi et al., 15 Oct 2025). The commonality is not the application domain but the central role of Poisson structure in the learned operator.

For imaging specifically, PoissonNet sits between two neighboring lines of work. Earlier deep Poisson denoisers, such as "Deep Convolutional Denoising of Low-Light Images" (Remez et al., 2017), treated photon-limited restoration directly in the image domain. More recent work such as "Moments Matter: Posterior Recovery in Poisson Denoising via Log-Networks" (Shoushtari et al., 8 Oct 2025) argues that standard MMSE denoisers recover only posterior means and proposes log-networks for posterior moment recovery. PoissonNet is different from both: it is not only a Poisson denoiser, and it is not a posterior estimator, but a Retinex-style decomposition network whose enhancement mechanism is explicitly regularized by Poisson noise statistics.

Its limitations are also specific. The model focuses on Poisson-dominated noise in extreme low-light conditions; the paper notes that real imaging pipelines often involve mixed noise, including Poisson noise, Gaussian read noise, and quantization noise, which the current formulation may not fully capture (Rao et al., 4 Jun 2025). It also acknowledges that when photon counts are extremely low, signal recovery remains fundamentally difficult even for Poisson-aware models. The future directions stated in the paper are accordingly pragmatic: handle mixed noise distributions more effectively and pursue more efficient solutions suitable for real-time applications (Rao et al., 4 Jun 2025).