Multiscale Sparsifying Transform Learning for Image Denoising

This repository contains the code associated with the following paper:

Ashkan Abbasi, Amirhassan Monadjemi, Leyuan Fang, Hossein Rabbani, Neda Noormohammadi, Yi Zhang, "Multiscale Sparsifying Transform Learning for Image Denoising," ArXiv Prepr., 2020.

Available Methods

The code for running the following methods are available in this package. All the codes needed to run the methods are included except for the sparsifying transform learning denoising.

• Four TLD (Sparsifying Transform Learning Denoising) based methods: TLD, MTLD, MMTLD, FMMTLD

  • Demo scripts: Benchmark_MTLD_for_Gaussian_denoising.m and Benchmark_MTLD_for_FMD.m
  • Requirements:
    • Download sparsifying transform learning [1-3] package (TSP2015ClosedformTL_code.zip) from here.
    • Extract the package in the METHODS folder. So, the path should be like this: ./METHODS/TSP2015ClosedformTL_code

• MM K-SVD (Multiscale Mixed K-SVD) and Fast MM K-SVD

  • Demo script: Benchmark_MMKSVD_for_Gaussian_denoising.m

• Fused K-SVD package contains K-SVD, MS K-SVD, and Fused K-SVD

  • Benchmark_FusedKSVD_for_Gaussian_denoising.m and Benchmark_FusedKSVD_for_FMD.m

• BLS-GSM

  • Benchmark_BLS_GSM_for_Gaussian_denoising.m
  • Instead of BLS-GSM, we use PURE-LET for denoising fluorescence microscopy images.

For PURE-LET, please refer to here.

All of the codes run over a Windows operating system with a proper MATLAB installation. However, just to let you know, we carried out our experiments using MATLAB R2019a over Windows 10.

Please, feel free to contact me or open an issue.

Datasets

(1) 12 Classic test images which are stored here.

(2) The image test_011 from BSD68 dataset.

(3) The test_mix subset of the Florescence Microscopy Denoising (FMD) dataset [1] (GitHub page, and Download link (Google Drive)).

Given a folder containing the images (e.g., ./DATASETS/20_classic_images), you can use Generate_synthetic_Gaussian_noise.m to generate noisy version of those images. For each given noise level, this script saves the noisy image and original image in separate .mat files and place them into a subfolder.

By storing noisy images in .mat files, we can ensure that our experiments are as repeatable as possible.

References

[1] Y. Zhang et al., “A Poisson-Gaussian Denoising Dataset With Real Fluorescence Microscopy Images,” in IEEE Conference on Computer Vision and Pattern Recognition, 2019, pp. 11702–11710.