search

cabouman / svmbir

Fast code for parallel or fan beam tomographic reconstruction
BSD 3-Clause "New" or "Revised" License
21 stars 8 forks source link

Blurry Demo Reconstructions #240

Closed XiaoWang-Github closed 2 years ago

XiaoWang-Github commented 2 years ago

Both the 2D and 3D Shepp Logan as well as microscopy demo reconstructions look blurry. See the attached images. Are the prior parameters set wrong?

3d_shepp_logan_recon 3D_microscopy_recon 2D_microscopy_recon

cabouman commented 2 years ago

Xiao, the microscopy recon is always going to be a little blurry due to the limited data. However, I agree that, on average, the recons tend to be a bit over regularized. This is closely related to the issue #239, which recommends reducing "regularization" for the proximal map mode.

I propose we try two changes in the prerelease branch that address both problems consistently. In auto_sigma_x we can change the scaling factor to 0.3 so that:

return_ 0.3 * auto_sigma_prior(sino, delta_channel, sharpness)

And in In auto_sigma_p we can change the scaling factor to 2.0 so that:

return 2.0 * auto_sigma_prior(sino, delta_channel, sharpness)

There could be some backward compatibility issues, but I think it's better to pull the bandaid off now. If this is agreeable to everyone, I can make the changes, and then you can try running the new code to see if you like it better.

XiaoWang-Github commented 2 years ago

Sounds good!

cabouman commented 2 years ago

I had an extensive discussion with Jordan about this, and I'm talking to different people to formulate the best solution. This is a very small change, but it needs to be done very carefully.

cabouman commented 2 years ago

OK, my solution to this problem was just to increase the value of sharpness in the demos. In the future, we may want to increase the default sharpness values in the code.