A Self-Supervised Blind-Spot Wavelet Residual Network for Label-Free Autofluorescent Image Denoising
Fairness Benchmarking for Self-Supervised Denoising
This repository provides a standardized framework for evaluating self-supervised denoising models. By using identical training budgets, datasets, and noise injections, this suite ensures a "fair" comparison between popular architectures and traditional algorithms.
Features
Unified Pipeline: Compare N2V, Neighbor2Neighbor, Self2Self, and WRTPNet in a single run.
Diverse Noise Profiles: Supports Gaussian, Poisson, and Poisson-Gaussian (mixed) noise.
Scientific Metrics: Automated calculation of PSNR, SSIM, MS-SSIM, FSIM, VIF, and FOM.
Statistical Validation: Generates Wilcoxon signed-rank tests to verify if improvements are statistically significant.
Reproducibility: Saves model weights (.pth) and experiment logs for every seed and noise type.
Installation
First, clone the repository and install the necessary dependencies:
git clone [https://github.com/Jacdencity/WRTPNet.git. cd denoising-fairness pip install -r requirements.txt
Dataset Structure
Place your clean images in a folder. The script will recursively find all images (PNG, JPG, BMP, TIF).
/data ├── train/ │ ├── image1.png │ └── image2.png └── test/ ├── sample1.png └── sample2.png
Usage
To run the full benchmark across all noise types and seeds defined in Config:
'''python main.py'''
Developed for the research community to promote transparent and reproducible denoising benchmarks.