detectorGapRemoval_Pix2Pix
A Pix2Pix generator that fills the gaps between the modules of tiled X-ray detectors, in projections and in sinograms. The network refines a linear interpolation of the gaps and leaves all valid pixels unchanged.
Code, an inference notebook and the training script: https://github.com/J-3TO/detectorGapRemoval_Pix2Pix
Files
| file | content | use |
|---|---|---|
pix2pix_projections_sinograms_generator.pth |
generator weights (69 MB) | inference |
pix2pix_projections_sinograms_full.pth |
generator + discriminator weights (80 MB) | continuing training (starts with fresh optimizers) |
Usage
# pip install "git+https://github.com/J-3TO/detectorGapRemoval_Pix2Pix"
from gap_pix2pix import inpaint, load_generator, pilatus3_2m_gap_mask
generator = load_generator("hf", device="cuda") # downloads the generator weights
filled = inpaint(projections, pilatus3_2m_gap_mask(), generator) # (N, H, W) or (H, W)
A complete example on simulated data: inference_example.ipynb in the GitHub repository.
Continue training from these weights in a clone of the GitHub repository: set the data paths
(the data format is described in its README) and init_weights: hf:pix2pix_projections_sinograms_full.pth
in config.yml, then run python3 train.py --config config.yml.
Input: flat-field corrected images and a gap mask made of complete rows and columns. The
per-image normalization and the linear pre-interpolation are done by inpaint.
Training
Trained on projections and sinograms of a lab nano-CT with a DECTRIS Pilatus3 2M detector, plus ImageNet validation images, with synthetic gaps of 7–40 pixels in 256 × 128 patches. Loss: adversarial (PatchGAN) + MSE + gap-mean loss + Earth Mover's Distance between soft histograms of the gap intensities. The released checkpoint (epoch 1199 of the last training stage) was selected by the total validation loss. Details: see the GitHub repository.
Limitations
Trained on data of one detector type and one setup; the behaviour on other detectors, geometries and photon statistics has not been tested.
License
The model weights are licensed under the GNU General Public License v3.0, the same licence as the code on GitHub. If you use them, please cite the paper below.
Citation
If you use this model, please cite:
TODO: add the reference of the paper.