# WIPUNet: A Physics-inspired Network with Weighted Inductive Biases for Image Denoising

arXiv: https://arxiv.org/abs/2509.05662 · HTML: https://mlsanity.com/paper/2509.05662/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 6 of its 14 references that are in the map.
- First posted on arXiv 6 Sep 2025.

## Builds on (most cited first, 6 of 6)

- [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://mlsanity.com/paper/1505.04597/) (2015)
- [Restormer: Efficient Transformer for High-Resolution Image Restoration](https://mlsanity.com/paper/2111.09881/) (2021)
- [Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising](https://mlsanity.com/paper/1608.03981/) (2016)
- [FFDNet: Toward a Fast and Flexible Solution for CNN based Image Denoising](https://mlsanity.com/paper/1710.04026/) (2017)
- [DnLUT: Ultra-Efficient Color Image Denoising via Channel-Aware Lookup Tables](https://mlsanity.com/paper/2503.15931/) (2025)
- [LAN: Learning to Adapt Noise for Image Denoising](https://mlsanity.com/paper/2412.10651/) (2024)

## Abstract (arXiv)

In high-energy particle physics, collider measurements are contaminated by "pileup", overlapping soft interactions that obscure the hard-scatter signal of interest. Dedicated subtraction strategies exploit physical priors such as conservation, locality, and isolation. Inspired by this analogy, we investigate how such principles can inform image denoising by embedding physics-guided inductive biases into neural architectures. This paper is a proof of concept: rather than targeting state-of-the-art (SOTA) benchmarks, we ask whether physics-inspired priors improve robustness under strong corruption. We introduce a hierarchy of PU-inspired denoisers: a residual CNN with conservation constraints, its Gaussian-noise variants, and the Weighted Inductive Pileup-physics-inspired U-Network for Denoising (WIPUNet), which integrates these ideas into a UNet backbone. On CIFAR-10 with Gaussian noise at σ∈15,25,50,75,100, PU-inspired CNNs are competitive with standard baselines, while WIPUNet shows a widening margin at higher noise. Complementary BSD500 experiments show the same trend, suggesting physics-inspired priors provide stability where purely data-driven models degrade. Our contributions are: (i) translating pileup-mitigation principles into modular inductive biases; (ii) integrating them into UNet; and (iii) demonstrating robustness gains at high noise without relying on heavy SOTA machinery.