Linear Neural Network as a Fast Solver for Dictionary Learning

This repository contains codes for the paper titled "Linear Neural Network as a Fast Solver for Dictionary Learning".

Requirements: Anaconda with Python 3.8, MATLAB 2022

Follow these instructions to run the codes:

1. Clone this respository.
2. Download the MATLAB toolboxes for K-SVD and OMP within the directory containing the downloaded codes. Compile the MATLAB toolboxes following the instructions available within them. (Source)
3. Move the codes within the folder "matlab_codes" to the folder of ksvd toolbox (ksvdbox13).
4. Create a conda environment using requirements.txt

Codes for three experiments in the paper are provided in three numbered directories. Run them in the following order from within the FastSolver_DictL directory:

Experiment 1: uniqueness (Appendix D)

Create data triplets

python -m 1_uniqueness.1_1_create_data_triplets

Perturb the dictionaries and compute sparse codes

python -m 1_uniqueness.1_2_perturbation

Experiment 2: convergence (Sec. 3.A)

Create data triplets Run generate_synthetic_dataset.m using MATLAB BL1: Online dictionary learning

python -m 2_convergence.2_1_odl

BL2: K-SVD

python -m 2_convergence.2_2_ksvd

Proposed FastSolver

python -m 2_convergence.2_3_fastsolver

Experiment 3: denoising

Create noisy data Run generate_noisy_data.m using MATLAB Learn dictionary using BL1: K-SVD

python -m 3_denoising.3_1_dictl_denoising_ksvd

Learn dictionary using proposed FastSolver

python -m 3_denoising.3_2_dictl_denoising_fastsolver

Denoise image using learnt dictionay Run image_denoise.m using MATLAB