At the moment only reductions over contiguous non-overlapping regions are supported, e.g. windowed averaging. From a signal processing perspective this is not the ideal method for downscaling -- it would be better to apply a gaussian filter to attenuate out-of-band frequencies and then downsample. However this is not possible as long as dask.array.coarsen is used to generate the downscaled array. Look into using dask.array.map_overlap or the gaussian smoothing functionality in dask-image to implement a gaussian pyramid.
cc @mkitti
At the moment only reductions over contiguous non-overlapping regions are supported, e.g. windowed averaging. From a signal processing perspective this is not the ideal method for downscaling -- it would be better to apply a gaussian filter to attenuate out-of-band frequencies and then downsample. However this is not possible as long as
dask.array.coarsenis used to generate the downscaled array. Look into usingdask.array.map_overlapor the gaussian smoothing functionality indask-imageto implement a gaussian pyramid.cc @mkitti