摘要
Zero-crossing of a derivative of Gaussian filter is a well-known edge location criterion. Examples are the Laplacian, the second derivative in the gradient direction (SDGD) and the sum of the Laplacian and SDGD (PLUS). Derivative operators can easily be implemented by convoluting the primitive image with a derivative of a Gaussian. Gaussian filter displaces the equipotential of half height inwards for convex edge and outwards for concave edges. A Difference-of-Gaussian (DoG) filter is similar to the Laplacian-of-Gaussian but with opposite sign and causes a convex edge shift inwards. This paper introduces the Multiple-of-Gaussian niters to reduce curvature-based location error. Using a linear combination of N Gaussians(N】2) with proper weights, the edge shifts can be reduced to 1/(2N-3) of the ones produced by a similar Laplacian-of-Gaussian filter.
Zero-crossing of a derivative of Gaussian filter is a well-known edge location criterion. Examples are the Laplacian, the second derivative in the gradient direction (SDGD) and the sum of the Laplacian and SDGD (PLUS). Derivative operators can easily be implemented by convoluting the primitive image with a derivative of a Gaussian. Gaussian filter displaces the equipotential of half height inwards for convex edge and outwards for concave edges. A Difference-of-Gaussian (DoG) filter is similar to the Laplacian-of-Gaussian but with opposite sign and causes a convex edge shift inwards. This paper introduces the Multiple-of-Gaussian niters to reduce curvature-based location error. Using a linear combination of N Gaussians(N>2) with proper weights, the edge shifts can be reduced to 1/(2N-3) of the ones produced by a similar Laplacian-of-Gaussian filter.
