Open
Description
To reduce the number of computations (multiplications) during convolution, keep only a configurable number of low frequency components of the data volume.
Convolution by the Gaussian 'optical transfer function' practically throws out high-frequency components of the volume. The higher the real-space sigma is, the larger the effect is.
This bandpass filter approach might also reduce the need for the eps parameter when inverse transform is required, though it already function well in terms of drastically diminishing noise amplification inherent to frequency space deconvolution.