Patch-based-Diffusion-Model-for-multiplicative-noise-removal

Abstract

Image restoration, particularly deblurring and denoising, remains a critical challenge in com- puter vision. This paper presents the Patch Diffusion Inverse Solver for Multiplicative Noise (PaDIS-MN), a novel approach leveraging score-based generative models to address inverse prob- lems involving multiplicative noise. Unlike traditional methods that often struggle with complex noise distributions, our method operates in the logarithmic domain, effectively transforming mul- tiplicative noise into additive noise, which is amenable to diffusion models. PaDIS-MN employs a U-Net based Score Network trained to predict noise in image patches. During inference, an itera- tive Ordinary Differential Equation (ODE) solver combines the denoising capabilities of the score model with a data consistency term, ensuring fidelity to the observed corrupted measurement. We demonstrate the efficacy of PaDIS-MN on a standard image dataset, showcasing significant improve- ments in image quality (PSNR, SSIM, LPIPS) compared to the corrupted inputs. Our patch-based approach enables efficient processing of high-resolution images, making PaDIS-MN a robust solution for real-world image degradation.

Dataset

We gratefully acknowledge the use of the Large-scale CelebFaces Attributes (CelebA) Dataset in this work.

Liu, Z., Luo, P., Wang, X., & Tang, X. (2015). Large-scale CelebFaces Attributes (CelebA) Dataset. The Chinese University of Hong Kong, Multimedia Laboratory. Available at: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

Environment

The whole project work is done in the google colab environment.

Training

Firstly ,load the dataset CelebA in your google colab environment .Then,modify the path of the images of the dataset in the DATA_DIR of the Training.py file in the google colab.This will mark the commencement of the training of the model.

Denoising and Reconstrucction

From the checkpoint directory ,get the path of the trained model and place it in the BESTMODELPATH in Denoising.py file .You may modify the path of the test image .Then,run the file in the google colab environment.This will produce a visualisation of ground truth ,corrupted image and denoised image produced by using your trained model.

Citation

1) Hu, J. (2024). PaDIS: Patch-based Diffusion Models [Source code]. GitHub. Retrieved from https://github.com/jasonhu4/PaDIS

2) Vuong, A. (2025). sdemultiplicativenoise_removal [Source code]. GitHub. Retrieved from https://github.com/anvuongb/sdemultiplicativenoise_removal

Acknowledgement

We'd like to acknowledge the following open-source projects which have been instrumental in this work:

PaDIS: Patch-based Diffusion Models We thank Jason Hu for developing and maintaining the PaDIS repository. This project's work on learning image priors through patch-based diffusion models was highly insightful. Hu, J. (2025). PaDIS: Patch-based Diffusion Models [Source code]. GitHub. Retrieved from https://github.com/jasonhu4/PaDIS

SDE Multiplicative Noise Removal We also extend our gratitude to An Vuong for the sdemultiplicativenoiseremoval repository, which provided valuable insights into stochastic differential equations for noise removal. Vuong, A. (2025). sdemultiplicativenoiseremoval [Source code]. GitHub. Retrieved from https://github.com/anvuongb/sdemultiplicativenoise_removal