demo.superres_mace¶
This demo illustrates the solution of a MACE problem using Mann iteration and the stacked operators F and G. The forward model is a subsampling operation, and the prior agent is a denoiser, with several different options.
First a clean image is subsampled, then white noise is added to produce noisy data. This is used to define a forward agent that updates to better fit the data.
In a classical Bayesian approach, this update has the form \(F(x) = x + c A^T (y - Ax)\), for a constant c. In some contexts, it’s useful to have a mismatched backprojector, which is equivalent to replacing \(A^T\) with an alternative matrix designed to promote better or faster reconstruction. As shown in a paper by Emma Reid, (in preparation), this is equivalent to using the standard back projector but changing the prior.
This demo provides the ability to explore mismatched backprojectors by changing the upsampling method used to define \(A^T\). It also provides the ability to change the relative weight of data-fitting and denoising by changing mu.
This demo uses parallel construction based on the MACE formulation and Mann iterations, while superres_pnp.py uses the standard Plug and Play method.
Functions
Illustrate MACE reconstruction on an image superresolution problem. |