The use of patches in image processing is clearly an instance
of the “divide and conquer” strategy: since it is admittedly too
difficult to formulate a global prior/model for an entire image, patchbased
approaches process patches thereof, and combine the processed
patches to obtain the processed image. The early patch-based methods
(namely the seminal non-local means–NLM–denoising method)
extract patches from the noisy image, then process/denoise them
independently (or maybe collaboratively, as in BM3D), and finally
return them to their original locations (averaging overlapping pixel
estimates). In this keynote presentation, I will overview the class of GM-based
patch-based approaches to image restoration and reconstruction.
After reviewing the first members of this family of methods,
which addressed only denoising under Gaussian noise, I will describe
more recent advances, namely: denoising under non-Gaussian noise;
use of class-adapted GM priors, i.e., tailored to specific image
classes (e.g., faces, fingerprints); addressing of problems other than
denoising (namely, deblurring, super-resolution, compressive image
reconstruction), by plugging GM-based denoisers in the loop of an
iterative algorithm (in what has recently been called the plug-and-play
approach); joint restoration/segmentation of images.