Poisson2Gaussian: Noise Gaussianization to Enhance Image Denoising

Modern image denoising often struggles with real-world noise, which is frequently dominated by Poisson-mixed noise (MPGN). Unlike additive white Gaussian noise (AWGN), MPGN is signal-dependent, heteroscedastic (variance changes with signal intensity), and statistically asymmetric. These characteristics make it difficult for standard deep learning denoisers to model and remove noise effectively. The authors investigate whether explicitly transforming this complex noise into a simpler, Gaussian-like distribution before denoising can improve performance.

The authors propose Poisson2Gaussian (P2G), a framework that treats noise removal as a two-stage process: noise Gaussianization followed by denoising. P2G uses an invertible normalizing flow—a type of neural network capable of learning complex probability distributions—to map the residual noise (the difference between the noisy and estimated clean image) to an i.i.d. Gaussian target. By using an exact likelihood objective, P2G aligns the full density of the noise rather than just matching low-order moments like variance. The framework is designed to be unbiased and architecture-agnostic, meaning it can be integrated with various existing denoising networks without requiring paired clean data.

Experiments demonstrate that P2G consistently outperforms existing methods across diverse datasets, particularly in low-light scenarios where Poisson noise is dominant. By reshaping the noise distribution, P2G allows standard denoisers to achieve higher peak signal-to-noise ratio (PSNR) values, with improvements of up to 0.75 dB in challenging cases. The authors show that the performance gain is robust across different network architectures, confirming that Gaussianization is a universal strategy for enhancing denoising quality.

This work provides a principled way to bridge the gap between idealized noise assumptions (AWGN) and the complex reality of sensor noise. By decoupling the task of noise characterization from the task of signal reconstruction, P2G offers a modular and highly effective tool for researchers and practitioners working in fields like low-light photography, microscopy, and astronomy, where high-quality denoising is essential for data analysis.