Abstract
Image noise can often be accurately fitted to a Poisson-Gaussian distribution. However, estimating the distribution parameters from only a noisy image is a challenging task. Here, we study the case when paired noisy and noise-free samples are available. No method is currently available to exploit the noise-free information, which holds the promise of achieving more accurate estimates. To fill this gap, we derive a novel, cumulant-based, approach for Poisson-Gaussian noise modeling from paired image samples. We show its improved performance over different baselines with special emphasis on MSE, effect of outliers, image dependence and bias, and additionally derive the log-likelihood function for further insight and discuss real-world applicability.
Introduction
In digital photography, noise is generally modeled as a combination of two factors: shot noise (Poisson distributed, varying with light intensity) and read noise (Gaussian distributed, constant across the sensor). Accurately estimating the parameters of this "Poisson-Gaussian" model is essential for designing effective denoising algorithms and analyzing sensor performance.
The traditional challenge is that we usually only have the noisy image to work with. Existing state-of-the-art methods, such as the seminal work by Foi et al., rely on complex local estimations of mean and variance from a single image. However, when researchers build new datasets or calibrate sensors, they often have access to paired samples—a noisy image and a corresponding "noise-free" version (often created by averaging many frames). Despite having this high-quality "ground truth" information, there has been no dedicated method to use it to achieve more precise noise modeling until now.
Methodology: The PoGaIN Approach
The proposed PoGaIN (Poisson-Gaussian Image Noise) method shifts the problem from local image analysis to a global statistical approach. The model assumes the observed signal $y$ is a function of the true signal $x$, a scaling factor $a$ (related to quantum efficiency), and a Gaussian component with variance $b^2$.
Instead of solving a computationally expensive Maximum Likelihood Estimation (MLE), PoGaIN utilizes cumulants. By deriving equations for the 2nd and 3rd cumulants of the noisy distribution in terms of the noise-free distribution, the authors create a system of two equations that can be solved directly for the unknown parameters $a$ and $b$. This global approach makes the method robust to complex textures where "flat" areas, required by local estimators, are hard to find. The use of the 3rd cumulant (related to skewness) is key to effectively "un-mixing" the Poisson and Gaussian components.
Experimental Results
The researchers evaluated PoGaIN against several baselines: the standard single-image method (FOI), a dedicated CNN, and a variance-based paired-data baseline (VAR). Testing was conducted on 62,500 image pairs synthesized from the BSD300 dataset.
Accuracy and Precision
PoGaIN (labeled "OURS" in evaluations) consistently achieved the lowest Mean Squared Error (MSE) for both noise parameters across a wide range of noise levels. Notably, for the read noise parameter ($b^2$), PoGaIN was approximately ten times more accurate than the VAR baseline.
Robustness and Image Independence
A key advantage of PoGaIN is its stability across different image contents. While deep learning baselines (CNN) were prone to wild misestimations on certain images, PoGaIN remained stable. This robustness stems from treating the entire image as a collection of samples rather than relying on local segmentation.
Bias and Theoretical Insights
The method also demonstrated superior bias characteristics. For read noise values above $b > 0.06$, PoGaIN showed virtually zero bias, whereas the variance-based method remained consistently biased. While the authors derived a log-likelihood function for theoretical insight, they found the cumulant-based approach to be far more computationally efficient while maintaining high accuracy.
Conclusion
PoGaIN provides a fast, accurate, and robust solution for noise modeling when paired samples are available. By leveraging the 3rd cumulant, it effectively separates shot and read noise components with minimal bias. Future work will focus on accounting for non-linear sensor behaviors like clipping and exploring hybrid models with optimized weighting schemes.
Summary updated