Zero-Shot Denoising via Neural Compression

This repository contains implementation of zero-shot denoising via neural compression : ZS-NCD

Initial Requirements

• Python > 3.11
• Jupyter Lab

Getting Started

Follow the steps below in the directory where you have cloned this GitHub repository:

1. Set up a Virtual Environment

You can choose either a conda environment or a venv-based environment.

Option A: Using Conda (Recommended)

If you have conda installed, create and activate a new environment:

conda create -n zs-ncd python=3.11
conda activate zs-ncd

Option B: Using venv

If you prefer not to use conda (already have python > 3.11 installed):

python3 -m venv venv
source venv/bin/activate

For Windows, activate the virtual environment with:

venv\Scripts\activate

2. Install Jupyter Lab

Option A: Using Conda

conda install -c conda-forge jupyterlab

Option B: Using pip

pip install jupyterlab

3. Launch Jupyter Lab

jupyter lab

4. Install Additional Packages

Once inside the notebook, follow the instructions in the first set of cells to install the required dependencies. (For a full list of dependencies, see requirements.txt).

---

Denoising via ZS-NCD

Enjoy experimenting! All instructions and examples can be found in the accompanying Jupyter notebook.

Zero-shot Denoisers at a Glance

For further discussion, refer to Figure 2 in the paper and the discussion therein.

Zero-shot denoising of Kodim05 with AWGN (σ = 25). Left: PSNR versus training iterations for zero-shot denoisers. Performance of BM3D and Restormer are included as a classical baseline and as a supervised empirical upper bound, respectively. Right: Visual reconstructions with PSNR/SSIM above each image. Compression-based denoising based on JPEG2K achieves inferior performance. Learning-based zero-shot denoisers often struggle with either overfitting or high bias. DIP and DD require early stopping to avoid overfitting. ZS-N2S and S2S struggle with high-resolution color images, and ZS-N2N often produces noisy outputs with potential overfitting. BM3D tends to oversmooth the denoised image. In contrast, ZS-NCD avoids these issues.