Speaker
Sebastian Neumayer
Description
In this talk, I will discuss various way to introduce spatial adaptivity into filter-based regularisation functionals. With this adaptivity, we are able to cancel out the filter responses to structure. Hence, we can interpret it as boosting the initial regulariser based on the data. A direct question if is we can repeat this process to get even better solutions. If we try this naively, the answer is sadly no for many cases. However, we can instead train the model in such a way that this procedure works, which opens up many interesting relations with other image reconstruction approaches. Our numerical results are on par with other approaches that rely on spatial adaptivity.
