Sobel¶
A first-order derivative filter that estimates the intensity gradient at each pixel using a pair of small (3×3) directional kernels, one for each axis. The combined gradient magnitude highlights edges — regions of rapid intensity change — cheaply and with reasonable noise robustness, since the kernel itself does a small amount of smoothing as part of computing the derivative.
See also
Scharr — a more rotationally accurate alternative to Sobel
Gaussian — smoothing that’s often applied before gradient filters
Laplace — the second-order counterpart to this first-order filter
Feature Extraction — every other feature