Imaging inverse problems arise in a wide range of applications, from medical to astronomical imaging.
Since they are ill-posed, they are usually formulated as regularized optimization problems.
I will first introduce classical variational approaches and their data-driven extensions, such as plug-and-play methods and learned regularizers.
The optimization algorithms used to solve these problems, however, scale poorly with the image size.
Multilevel methods, inspired by multigrid solvers for PDEs, address this issue: they exploit a multiscale decomposition of the image, e.g. through the wavelet transform, to define auxiliary problems at coarser scales whose solutions accelerate convergence on the original problem.
Because coarse levels mainly capture low frequencies, these schemes are efficient when the degradation is strong and the noise is low, but lose their advantage otherwise.
Building on the connection between multilevel and block-coordinate descent algorithms, we propose a strategy that incorporates high-frequency updates into multilevel schemes.
It relies on an automatic block selection rule that adapts to the degradation and noise levels, accelerating convergence across all degradation regimes.