The repository contains the implementatiion of the paper Non-linear Aggregation of Filters to Improve Image Denoising. This project is a partial requirement of the Digital Image Processing course at IIIT Hyderabad instructed by Prof. Ravi Kiran Sarvadevabhatla in the Monsoon 2020 semester.
The paper introduces a novel aggregation method to efficiently perform image denoising. Preliminary filters are aggregated in a non-linear fashion, using a new metric of pixel proximity based on how the pool of filters reaches a consensus. It provides a theoretical bound to support our aggregation scheme, its numerical performance is illustrated and we show that the aggregate significantly outperforms each of the preliminary filters.
git clone https://github.com/Digital-Image-Processing-IIITH/project-revision.gitconda create --name envname python=3.8
conda activate envnameEnsure that you install all the dependencies in the virtual environment before running the program. We have used Python 3.8 during the development process. Do ensure that you have the same version before running the code.
cd src
python3 main.pyDo note that the training process may take several hours. The team members used the Ada High Performance Cluster of IIIT Hyderabad for training the model.
The link to the best model obtained upon training is here.
To use the pretrained model download the pkl file to src/ and in src/main.py set:
loadModel = True
The following command will install the packages according to the configuration file src/req.txt.
pip3 install -r src/req.txt- cycler==0.10.0
- joblib==0.17.0
- kiwisolver==1.3.1
- matplotlib==3.3.3
- numpy==1.19.4
- pandas==1.1.4
- Pillow==8.0.1
- pycobra==0.2.3
- pyparsing==2.4.7
- python-dateutil==2.8.1
- pytz==2020.4
- scikit-learn==0.23.2
- scipy==1.5.4
- seaborn==0.11.0
- six==1.15.0
- threadpoolctl==2.1.0
project-revision
├── dataset
| ├── test
| ├── train
├── docs
├── misc
| ├── dataset
| ├── reszie.py
├── src
| ├── denoise
| | |── __init__.py
| | |── denoise.py
| | |── errors.py
| ├── noise
| | |── __init__.py
| | |── errors.py
| | |── noise.py
| ├── results
| ├── cobra.py
| ├── cobramachine.py
| ├── denoise.py
| ├── helper.py
| ├── main.py
| ├── req.txt
├── README.md
├── guidelines.md
├── proposal.md
The dataset folder contains the train and test splits of the dataset. The docs folder contains the project proposal documents, and project presentation slide-decks.The results folder holds the results of running the implementation on the images. The src folder contains the source code and the dependent libraries.
The helper.py has the denoising evaluation function and the training data loader function. The cobra.py and cobramachine.py files have the COBRA algorithm defined in the Python library pycobra. denoise/denoise.py and noise/noise.py files contains all the denoising algoithms and noise models respectively. main.py is the starting point of the software execution. It calls all the required functions in the appropriate order and stores the final model.
The dataset is made of up 26 images taken from the Labeled Faces in the Wild public benchmark dataset for face verification. The images were resized to 64 x 64. The implementation considers 25 images in the training set, and takes one image for testing purpose.
The program takes 25 input noise-free images and adds seven standard noise models to it on the run-time. The noisy images are run through seven classical denoising algorithms. We train the COBRA machine using the noisy and denoised images to perform non-linear aggregation of the preliminary filters. The best model obtained from the training process is used for image denoising.
 Amogh Tiwari |
 Dolton Fernandes |
 George Tom |
 Naren Akash R J |
All the team members are undergraduate research students at the Center for Visual Information Technology, IIIT Hyderabad, India.
The team members are grateful towards the instructor Prof. Ravi Kiran Sarvadevabhatla and mentor teaching assistant Soumyasis Gun. They also acknowledge CVIT, IIIT Hyderabad for providing access to the Ada High Performance Cluster for training the machine learning model.
The software can only be used for personal/research/non-commercial purposes. To cite the original paper:
@InProceedings{10.1007/978-3-030-52246-9_22,
author="Guedj, Benjamin and Rengot, Juliette",
editor="Arai, Kohei and Kapoor, Supriya and Bhatia, Rahul",
title="Non-linear Aggregation of Filters to Improve Image Denoising",
booktitle="Intelligent Computing",
year="2020",
publisher="Springer International Publishing",
address="Cham",
pages="314--327",
abstract="We introduce a novel aggregation method to efficiently perform image denoising. Preliminary filters are aggregated in a non-linear fashion, using a new metric of pixel proximity based on how the pool of filters reaches a consensus. We provide a theoretical bound to support our aggregation scheme, its numerical performance is illustrated and we show that the aggregate significantly outperforms each of the preliminary filters.",
isbn="978-3-030-52246-9"
}