Introduction

This repository contains implementations and pre-trained models for various image restoration tasks including denoising and deblurring. The models have been tested on multiple datasets and configurations to evaluate their performance.

• Denoising Models: REDNet, DnCNN, Restormer, MaIR
• Deblurring Models: DeblurGANv2, Restormer, MaIR

Furthermore, there is a Gradio-based web demo available for interactive testing of the models.

Pre-requisites

• Python 3.11+
• GNU Make (optional, for downloading weights and datasets)
• CUDA-capable GPU (optional but recommended for faster inference and compatibility with pre-trained weights)

Set up

Installation

make install-packages

Or manually install the required packages:

pip install uv
uv pip install -r requirements.txt \
    torch==2.7 torchvision --extra-index-url https://download.pytorch.org/whl/cu126 \
    https://github.com/state-spaces/mamba/releases/download/v2.2.5/mamba_ssm-2.2.5+cu12torch2.7cxx11abiTRUE-cp311-cp311-linux_x86_64.whl

Note:

• Use uv for better and faster dependency resolution.
• The torch version should match mamba's supported versions. Adjust the CUDA version as needed.
• Find the pre-built wheels for mamba here.

Download Pre-trained Weights

# ~11.6 GB
make download-weights

Download Datasets (Optional)

This is only required for running the full test suite or if you want to run the demo app with these datasets.

# ~2 GB
make download-datasets

Run Tests

Full test suite

python scripts/tests.py

This will run tests for all models and tasks on the full datasets. All images and results are stored in the results/ directory. The test configurations are detailed below.

Quick tests

Small-scale tests for demo purposes:

python scripts/test_demo.py

This will run tests for all models and tasks on a single image from representative datasets. All images and results are stored in the demo/ directory.

Test Configurations

These tests cover a variety of image restoration tasks including denoising and deblurring using different datasets and models. After running the tests, the results will be saved in results/ directory and a summary CSV file results/results_summary.csv.

Denoising

Gaussian Noise

• Gray Image
  • Non-blind
    • Sigmas: 15, 25, 50
    • Datasets: Set12, BSD68, Urban100
    • Models: REDNet (sig=50 only), DnCNN, Restormer
  • Blind
    • Sigmas: 15, 25, 50
    • Datasets: Set12, BSD68, Urban100
    • Models: DnCNN, Restormer
• Color Image
  • Non-blind
    • Sigmas: 15, 25, 50
    • Datasets: CBSD68, Kodak, McMaster, Urban100
    • Models: Restormer, MaIR
  • Blind
    • Sigmas: 15, 25, 50
    • Datasets: CBSD68, Kodak, McMaster, Urban100
    • Models: DnCNN, Restormer

Real Noise

• Datasets: SIDD
• Models: Restormer, MaIR

Deblurring

Defocus Blur

• Datasets: DPDD
• Models: Restormer (single-image, dual-pixel)

Motion Blur

• Datasets: GoPro, HIDE, RealBlur_J, RealBlur_R
• Models: DeblurGANv2 (fpn_inception, fpn_mobilenet), Restormer, MaIR

Web Demo

To run the Gradio-based web demo for interactive testing of the models:

gradio scripts/demo.py

Test Results

Denoising

Deblurring

Note:

• Avg. time: Average inference time per image (in seconds) on an NVIDIA Tesla T4 GPU.
• GFLOPs: Number of floating-point operations (in billions) for processing a 256x256 image.

Visualization

Denoising

Deblurring

Credits

• The original implementations of the models used in this repository are credited to their respective authors:

  • REDNet
  • DnCNN
  • DeblurGANv2
  • Restormer
  • MaIR

• The tool for converting Caffe models (used in REDNet) to PyTorch is credited to caffemodel2pytorch.

• The datasets used for testing are provided by Restormer repository.