Regularization with external denoiser

[ozanoktem]

Now that Holger has implemented ADMM (issue #1198 and issues #1201 and #1202), it could be worthwhile to check out Regularization by Denoising (RED). This is an approach that is close to task based reconstruction that we are pursuing, but for denoising only.

RED offers a plug-and-play scheme that is capable of incorporating any image denoising algorithm, typically given by machine learning, and can treat general inverse problems. It corresponds to learning a prior that incorporates the denoising and the approach relies heavily on the ADMM-type of optimization technique in order to handle denoising and reconstruction in an intertwined manner. Since it solves an optimisation problem the authors can provide converge.

[kohr-h]

and the approach relies heavily on the ADMM-type of optimization technique

Is that true? I read the abstract as if they don't need that in contrast to earlier work. Looks interesting in any case.

[ozanoktem]

@kohr-h is right, I was sloppy reading the abstract. The Plug-and-Play Prior (P^3) method is connected to ADMM, but the RED method seems to be compatible with a wider range of optimisation solvers, including gradient descent, ADMM, and a 'fixed-point method'.