danaroth/unrolled_unet_demosaicing

Description

Series of weights recovered by training the ligthning torch model UnrolledSystem implemented in: (Unrolled demosaicking)[https://github.com/mattmull42/unrolled_demosaicking]

The network was trained over 15 color filter array patterns:

• bayer
• binning
• chakrabarti
• gindele
• hamilton
• honda
• honda2
• kaizu
• kodak
• luo
• quad_bayer
• random
• sparse_3
• wang

The network is based on U-NET, unrolled over 4 stages, and plugged into an ADMM solver. The network is trained over 300 natural images, cut into patches of size 64 x 64.

The three versions given in this repository are:

• 4: Baseline weights.
• 4B: Variant with different training set.
• 4V: Introduces geometric transformations on the patterns.

Citation

If you use this dataset, please cite:

@InProceedings{muller_eusipco_2024,
  author    = {Muller, Matthieu and Picone, Daniele and Dalla Mura, Mauro and Ulfarsson, Magnus Orn},
  booktitle = {European Signal Processing Conference ({EUSIPCO})},
  title     = {Pattern-invariant unrolling for robust demosaicking},
  year      = {2024},
}