Deep Learning-Assisted Dynamic Mode Decomposition (DA-DMD) for NRB removal in CARS Spectroscopy

DMD decomposes noisy input (CARS spectra) into different modes based on their frequency. Deep learning part uses SE Block for channel atention to weigh relevance of modes and then a CNN Block to extract the final clean output (Raman spectra). The noise (Non-resonant Background) has low frequency while the clean spectra (Raman signatures) have higher frequencies. This criteria makes the background removal possible. To know more (link to paper).

DA-DMD method.

Usage

Setup the project:

• Step 1: Clone this repository and create a virtual environment to isolate the dependencies.

# Clone repo with default name DA_DMD
git clone https://github.com/spectra-analysis/DA_DMD.git
cd DA_DMD

# Create and activate a conda environment
conda create -n "env_dadmd" python=3.10
conda activate env_dadmd

• Step 2: Install required packages as given in requirements.txt, preferably with the same version.

pip install -r requirements.txt

Get started: Understand the code usage with example_dadmd.ipynb notebook that illustrates a training and testing example.

Training: You may train a new DA-DMD model using train_dadmd.py. You may use the given synthetic CARS-Raman data pair or use synthetic generator 1, 2 or 3.

Testing: You may test the trained model using test_dadmd.py.

Note: We did our experiments on Ubuntu 22.04 with Python 3.10, PyTorch 2.2.1 and CUDA 12.1.

Citation

Authors: Adithya Ashok Chalain Valapil, Carl Messerschmidt, Maha Shadaydeh, Michael Schmitt, Jürgen Popp, Joachim Denzler.

Publisher: DAGM German Conference on Pattern Recognition (DAGM-GCPR) 2025

BibTeX:

@inproceedings{valapil2025dadmd,
  title     = {Deep Learning-Assisted Dynamic Mode Decomposition for NRB removal in CARS Spectroscopy},
  author    = {Adithya Ashok {Chalain Valapil} and Carl Messerschmidt and Maha Shadaydeh and Michael Schmitt and Jürgen Popp and Joachim Denzler},
  booktitle = {DAGM German Conference on Pattern Recognition (DAGM-GCPR)},
  year      = {2025}
}