BlurProof is a deep learning framework for radio interferometric image deconvolution. It generates realistic dirty images by convolving clean astronomical images with physically accurate PSFs derived from real interferometer configurations (ALMA, VLA, ngVLA). The model learns to reconstruct the original clean image, effectively learning the inverse of the imaging process.
Note: The model architecture, data preprocessing, and PSF generation pipeline are fully implemented. However, the model itself has not been trained yet — all scripts and configurations are in place and ready for training.
In radio interferometry, incomplete sampling of the spatial frequency (u–v) plane produces a dirty image: [ I_\text{dirty} = I_\text{true} * B_\text{dirty} ] where (B_\text{dirty}) is the point spread function. BlurProof replaces traditional iterative deconvolution (like CLEAN) with a learned U-Net model trained on synthetic data.
The clean images come from the Galaxy Zoo 2 dataset, and the PSFs are generated from realistic interferometer configurations.
BlurProof/
├── train.py # Model training entry point
├── array_configurations/ # ALMA / ngVLA / VLA cfg files
└── utilities/ # PSF generation and visualization tools
git clone https://github.com/harry353/BlurProof.git
cd BlurProof
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtGenerate PSFs:
python utilities/create_psf.pyInspect samples:
python utilities/inspect_sample.pyTrain the model:
python train.pyReleased under the MIT License.