Discrete approximations of Gaussian smoothing and Gaussian derivatives
The research compares three ways to discretize Gaussian smoothing and its derivatives for image data: sampling kernels, local integration, and using a discrete analogue of the Gaussian kernel with central difference operators. Findings show that the discrete analogue is theoretically superior at fine scales, while sampled kernels are better for large scales. The choice depends on whether fine or coarse scale accuracy is prioritized.