CSAS Deconvolution

Code for SINR: Deconvolving Circular SAS Images Using Implicit Neural Representations found at https://arxiv.org/abs/2204.10428.

Requirements

1. Python >= 3.6
2. Conda

   Installation

3. Update conda using conda update -n base conda.
4. Install the python virtual environment containing the necessary packages using conda env create -f environ.yml.
5. After installing, enter the virtual environment with conda activate CSAS_INR_Deconv.
6. Finally, install PyTorch using from command line like conda install pytorch torchvision cudatoolkit=10.2 -c pytorch. Visit https://pytorch.org/ for more details.

Deconvolving Simulated data

The example_sim_deconv_pipeline/deconvolve_simulated_scene.py script creates two sets of simulated CSAS measurements, reconstructs the measurements using DAS, and then deconvolves the images using our INR approach and baselines discussed in the paper. The simulation and geometry parameters are modified by editing using the example_sim_deconv_pipeline/system_parameters.ini and example_sim_deconv_pipeline/simulation.ini files. The deconvolution parameters can be edited using the example_sim_deconv_pipeline/deconv.ini. All .ini files are commented to provide instructions for use. The output of the deconvolution methods are saved in the deconv_dir directory.

Running the pipeline will generate a results figure in each deconv_dir/image* directory showing the deconvolution results of all the methods run on the created dataset.

Additionally, one can find bar plots displaying deconvolution metrics for each method.

Deconvolving AirSAS data

The script airsas_deconv_pipeline/reconstruct_scene.py will reconstruct an AirSAS scene, compute the PSF, and then deconvolve the scene using the methods specified in the associated deconv.ini file. For example, one can use the INR to deconvolve AirSAS small feature cutout scene shown in the paper figure:

.

Running airsas_deconv_pipeline/reconstruct_scene.py will populate the directory airsas_deconv_pipeline\20k_scene with INR deconvolution results for the 20k small feature DAS scene.

Running the script shoould yield a DAS reconstruction that looks like this:

DAS Reconstruction (found at airsas_deconv_pipeline\20k_scene\scene_abs.png) :

and an INR Deconvolution (found at airsas_deconv_pipeline\20k_scene\image0\INR\deconv_img_100.png) at epoch 100 that looks like this: