Standard image processing frameworks typically treat light as a continuous field. This assumption justifies the ubiquitous use of the Mean Squared Error (MSE) loss function, which is mathematically derived from a Gaussian noise model. However, in photon-starved regimes—such as nanoscale imaging, low-dose medical tomography, or long-range Lidar—light reveals its discrete quantum nature. Instead of a continuous signal, sensors record the arrival of individual, sparse photons.
As noted by Kaganovsky et al. in their work on transmission tomography, standard Gaussian noise approximations break down entirely when photon counts are low. In these environments, the measurement process is governed by Poisson statistics, where the noise is fundamentally signal-dependent (heteroscedastic): the variance of the noise is exactly equal to the signal intensity itself ($\sigma^2 = \mu$). Directly modeling this Poisson likelihood, rather than relying on Gaussian approximations, is the key to recovering high-fidelity signals from extreme noise.
Why Gaussian Approximations and Transformations Fail
To understand why direct Poisson modeling works, we must first look at why traditional methods fail. The standard MSE loss function assumes homoscedasticity—meaning the noise variance is uniform across the entire image, regardless of pixel brightness. Under Poisson noise, however, bright regions exhibit high absolute variance, while dark regions exhibit low absolute variance. Applying a standard quadratic loss treats all absolute deviations equally, causing optimization algorithms to over-smooth faint structural details or severely misestimate the true signal flux.
A common classical workaround is the Variance-Stabilizing Transformation (VST), such as the Anscombe transform. A VST attempts to mathematically map Poisson-distributed data into a Gaussian space so that standard denoising algorithms can be applied. However, Giryes et al. demonstrate in Sparsity Based Poisson Denoising with Dictionary Learning that VSTs fail catastrophically in extreme low-light conditions where photon counts drop near zero. When data is highly sparse, the discrete, grainy nature of the signal cannot be smoothed into a continuous Gaussian curve without destroying the underlying structural information.
The Intuition of the Poisson Log-Likelihood Loss
Rather than transforming the data to fit a mismatched noise model, modern computational imaging frameworks embed the physical statistics of photon arrival directly into the objective function. The probability of observing $y$ photons given an underlying true intensity $x$ is defined by the Poisson distribution:
$$P(y \mid x) = \frac{x^y e^{-x}}{y!}$$
Taking the negative logarithm and dropping the constant factorial term yields the physics-informed Poisson log-likelihood loss:
$$L(x; y) = x - y \log x$$
The core intuition of why this loss function outperforms MSE lies in its gradient. For standard MSE, the loss is $\frac{1}{2}(x - y)^2$, and its gradient with respect to the prediction $x$ is simply the linear residual:
$$\frac{\partial L_{\text{MSE}}}{\partial x} = x - y$$
In contrast, the gradient of the Poisson log-likelihood loss is:
$$\frac{\partial L_{\text{Poisson}}}{\partial x} = 1 - \frac{y}{x} = \frac{x - y}{x}$$
The denominator $x$ changes the optimization dynamics completely. The error $(x - y)$ is adaptively scaled by the expected intensity $x$. If the algorithm predicts a very dim region (small $x$), even a tiny discrepancy between the prediction and the observation results in a massive gradient step, heavily penalizing the error. If the algorithm predicts a bright region (large $x$), the penalty is naturally relaxed to accommodate the higher inherent physical variance of shot noise.
As shown by Aich et al. and Bui et al., integrating this term directly into the data fidelity component ensures that the optimization process respects the true physical relationship between signal intensity and noise variance across varying intensity regions.
Regularization and Optimization in Sparse Regimes
Minimizing the Poisson log-likelihood alone is often an ill-posed problem, especially when data contains missing regions or extreme noise. To stabilize the inversion, the Poisson loss is paired with structural priors.
In classical sparse recovery, the objective combines the negative Poisson log-likelihood with a sparsity-promoting penalty (such as the $\ell_1$-norm or Total Variation). Rohban et al. proved that an $\ell_1$-constrained Maximum Likelihood estimator achieves minimax optimality for Poisson statistics, fundamentally outperforming Gaussian-based methods like the rescaled LASSO. Algorithmic frameworks like SPIRAL-TAP, developed by Harmany et al. in This is SPIRAL-TAP: Sparse Poisson Intensity Reconstruction ALgorithms - Theory and Practice, leverage these penalized likelihood formulations to iteratively reconstruct signals while strictly enforcing non-negativity constraints.
Recently, this physics-informed loss has been integrated into deep learning architectures. For instance, Lawrence et al. combined a Poisson likelihood objective with an untrained neural network prior (the Deep Decoder) in Phase Retrieval with Holography and Untrained Priors: Tackling the Challenges of Low-Photon Nanoscale Imaging. This approach successfully reconstructed high-quality nanoscale images from just 1 photon per pixel, bypassing the need for large training datasets. Similarly, Monroy et al. demonstrated that self-supervised denoising models trained with negative log-likelihood loss can learn to extract underlying signals without ever seeing a clean ground-truth image.
Adapting to Alternative Photon Statistics
Depending on the hardware configuration and imaging modality, the underlying Poisson process manifests in different statistical variations that require tailored loss functions:
- Single-Photon Counting (Bernoulli Statistics): When sensors like Single-Photon Avalanche Diodes (SPADs) operate in a binary mode—recording only whether at least one photon arrived during an ultra-short exposure window—the statistics shift from a Poisson distribution to a Bernoulli distribution. As Liu et al. and Hu et al. demonstrate, utilizing a Bernoulli-based logarithmic data fidelity term preserves critical statistical information that is typically lost when binary frames are simply summed into a single image, drastically improving reconstruction fidelity in high-contrast or ultra-low-signal environments.
- Image Differencing (Skellam Statistics): In applications like background subtraction or change detection, algorithms must analyze the difference between two independent, noisy frames. Because the difference between two independent Poisson variables does not follow a Gaussian distribution, Hwang et al. utilized the exact Skellam distribution. This exact modeling prevents false alarms and preserves sharp object boundaries under fluctuating, low-light conditions.
By aligning the mathematical objective function with the true quantum statistics of light detection, these methods ensure that every captured photon contributes optimally to the final reconstruction.