Abstract
Some fast approximate methods have been developed for bilateral niters. One of them decomposes filter coefficients spatially and another one reduces computational costs by downsampling of voxel data in coordinate-graylevel product spaces. In this paper, we combine these two techniques to reduce computational times of bilateral filtering by spatially decomposing filter coefficients in linear Gaussian filters in the coordinate-graylevel product space and reducing the voxel sites by their downsampling. Though our method is slightly slower than the previous method of directly decomposing filter coefficients in image planes, approximation errors are smaller in our method than the previous technique.
