Paper analyzis

[balthazarneveu]

Papers

https://hal.univ-lorraine.fr/IDS/hal-03940525v1 https://hal.science/hal-03186499/file/papier_SSVM%20%281%29.pdf

Line artifacts on the butterfly.

A lead to tackle this problem would be to complete our database with patches generated from a sinusoı̈dal basis, as was done in [29]. -> why not simply extending primitives..

Natural Color sampling improves results :warning: Maybe needs harder patches to improve! (low color contrast for instance)

:warning: No mention of quantization and jpg compression (natural input images are most probably 8bits sRGB) :mag: to be checked! ... Deadleaves targets in float naturally have a different distribution. :question: Did they save the deadleaves to disk? or did they generate it on the fly?

[balthazarneveu]

May be worth a read:

• Unsupervised deblurring : https://arxiv.org/pdf/2210.05361.pdf
• Polyblur

[balthazarneveu]

Proposal for first meeting: We want to test 3 things: 1/ synthetic charts 2/ nafnet 3/ blind deblurring

Would be good to start with several tracks

1 + 2 to avoid mixing too much things at first e.g. => Blind denoising on synthetic charts + how much is NAFNet better?

• Blind denoising problem + Synthetic charts
• Stacked ConvNet Relu VS StackedConv + NAF multiplier
• same with a UNET Relu VS UNET + NAF multiplier

[balthazarneveu]

NAFNet

• U-Net like architecture Replace "conv+Relu+batch norm" base blocks by:
• layer norm
• pointwise 1x1 conv + 3x3 depthwise conv
• simple gate = chunk and multiply
• simplified channel attention

Note: Layer norm = normalize over the channels and spatial dimensions 💡 share very rough global image information at all levels)

[balthazarneveu]

NAFNet deblur training, change to deadleaves.

Removing camera blur , delbracio

https://arxiv.org/pdf/1505.02731.pdf