Transforming Noise Distributions with Histogram Matching: Towards a Single Denoiser for All

TL;DR

This work tackles the generalization gap of supervised Gaussian denoisers to unseen noise by introducing histogram matching to convert arbitrary noise into a target Gaussian form with known intensity $\sigma_0$, coupled with an iterative denoising cycle. The method combines global/local/frequency-domain histogram matching, intrapatch permutation, and pixel-shuffle down-sampling to disrupt spatial and channel correlations, followed by fixed-level and flexible Gaussian denoising and texture restoration to progressively refine the transformation. Across synthetic and real-world noises, including Poisson, salt-and-pepper, repeating patterns, and low-light real-world noise, the approach substantially improves PSNR and SSIM over strong baselines, enabling a single Gaussian denoiser to generalize to out-of-distribution noise. The results demonstrate practical impact: improved robustness and denoising quality without retraining on every new noise type, aided by an explicit cycle between noise transformation and denoising and by texture-aware refinements.

Abstract

Supervised Gaussian denoisers exhibit limited generalization when confronted with out-of-distribution noise, due to the diverse distributional characteristics of different noise types. To bridge this gap, we propose a histogram matching approach that transforms arbitrary noise towards a target Gaussian distribution with known intensity. Moreover, a mutually reinforcing cycle is established between noise transformation and subsequent denoising. This cycle progressively refines the noise to be converted, making it approximate the real noise, thereby enhancing the noise transformation effect and further improving the denoising performance. We tackle specific noise complexities: local histogram matching handles signal-dependent noise, intrapatch permutation processes channel-related noise, and frequency-domain histogram matching coupled with pixel-shuffle down-sampling breaks spatial correlation. By applying these transformations, a single Gaussian denoiser gains remarkable capability to handle various out-of-distribution noises, including synthetic noises such as Poisson, salt-and-pepper and repeating pattern noises, as well as complex real-world noises. Extensive experiments demonstrate the superior generalization and effectiveness of our method.

Figures (14)

• Figure 1: Visualization of our noise transformation effect. The first row is random impulse noise and the second row is real-world noise. DMID DMID is a diffusion-based powerful Gaussian denoiser, but it shows difficulty in removing out-of-distribution noise. Through our noise transformation, its denoising effect has been greatly improved.
• Figure 2: Overall flowchart of our method. The initial noise for transformation is obtained by subtracting the smoothed image from the original noisy image. Subsequently, appropriate noise transformation and shuffling strategies will be selected based on the signal, spatial and channel correlations of the noise. We then employ a fixed-level Gaussian denoiser to obtain the denoising result, which will be used for the next round of iteration after certain processing, forming a mutually reinforcing cycle.
• Figure 3: Qualitative comparison on synthetic noise. The noise images from the first row to the last row are respectively Bernoulli noise, speckle noise and circular repeating pattern noise images. They are all out-of-distribution noises that Gaussian denoisers are difficult to remove.
• Figure 4: Qualitative comparison on real-world noise. Real-world noise is much more complex than synthetic noise. Gaussian denoisers generally have difficulty removing real-world noise. But through our noise transformation, they have achieved excellent denoising results.
• Figure 5: Qualitative analysis of each step in our method on real-world noise.