danaroth/adapters

Description

This repository contains a series of weights for adapting the DRUNet denoiser in order to be able to work with hyperspectral images.

These weights are meant to be used with the hypnp library:

http://github.com/Danaroth83/hypnp

In particular the weights contained in this folder are associated to the following adapting architecture:

• 20251025_174116_606035 - projection_encoder: An encoding/decoder network.
• 20251029_110345_695678 - projection_qr: A QR decomposition encoder and deep decoder network.
• 20251026_035736_830488 - grouper_arranger: A band selection module.
• 20251026_044313_346452 - grouper_arranger_skip: A band selection module with skip attention network.
• 20251026_141518_492696 - film_middle: A FiLM that hooks middle layers of the DRUNet. Baseline result.
• 20251026_063231_578169 - film_middle_qr: FiLM network with QR projection of the input.
• 20251027_091506_635645 - film_middle_qr_groups_10: FiLM network with QR projection, with inputs passed sequentially in groups of 10.
• 20251029_093111_154168 - film_no_head: FiLM network without trained head in DRUNet.
• 20251031_173527_494979 - film_full: FiLM full network
• 20251102_135320_481765 - lora_pca_big: LoRA network applied on a PCA decomposition.
• 20251104_192334_965269 - cave_projection_matrix_orthogonal: Linear channel projection, with orthogonal matrix.
• 20251105_024251_967712 - cave_projection_matrix_orthogonal_compressed: Linear channel projection matrix to 3 channels. Matrix is constrained to be orthogonal.
• 20251104_200826_195587 - cave_projection_matrix_simplex: Linear channel projection matrix constrained to be positive with sum-to-one condition.
• 20251104_230322_748168 - cave_projection_matrix_simplex_compressed: Linear channel projection matrix to 3 channels. Matrix is constrained to be positive with sum-to-one condition.

Credits

These weights were produced by:

Daniele Picone
Univ. Grenoble Alpes, CNRS, Grenoble INP, GIPSA-lab, 38000 Grenoble, France
Mail: [email protected]

Mohamad Jouni
Univ. Grenoble Alpes, CNRS, Grenoble INP, GIPSA-lab, 38000 Grenoble, France
Mail: [email protected]