Hi Jerry
First a clarification. Yes, deconvolution "might" be used as a noise reduction tool, but it has not been designed with that in mind. Deconvolution procedures, apart from a naive filter inversion, includes what is called a "regularizator". The goal of these "objects" (or algorithms) is to prevent noise amplification in the procedure of recovering the true signal at high frequencies (or small scales). In the most simple regularizators, this means just to blur the image a little bit, and in most advanced ones some prior knowledge is used, like the image should be smooth with some sharp edges, or that the image is mostly black with a few bright dots.
So, the bottom line is that even when deconvolutions may be used as noise reduction algorithms (and we are implementing a new deconvolution also based in the total generalized variation), they are not recomended for this task. Even when they noise reduction algorithm share the same regularizator (as will be the case of tgvdenoise and tgvrestoration), noise reduction algorithm are optimized for this task, and include some ad-hoc routines to enhance the performance.
Now, about noise reduction algorithm guidelines... In my opinion TGVDenoise seems to be the better choise in most cases, for linear or non-linear data. The algorithm assumes that there is gaussian, homogeneus noise in the image, but with our local support implementation there is an extra flexibility that allows the algorithm to work quite well in presence of poisson noise, and other structurated noises.
If the image contains clearly distinguisable noise patterns, sinosoidal specially, then noise reduction in the fourier space is recommended. For that, I wrote a module called NotchFilter, similar to the DefectMap process, that works specially well. Also the TGVRestoration process could be used, in the fourier map mode that I'm including.
For pepper and salt noise, other non-linear filters like a median filter are more better suited.
And that's it, in a nutshell. There is still room for ATWT and MMT, specially with linear data or weird noise, but ACDNR is basically replaced by TGVDenoise. Also I won't take out SCNR, specially to fix color casts after noise reduction with other algorithms.
